<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>What I've Learned on GABE | Claude Team Setup for Small Teams</title><link>https://gabezen.com/blog/</link><description>Recent content in What I've Learned on GABE | Claude Team Setup for Small Teams</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 01 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://gabezen.com/blog/index.xml" rel="self" type="application/rss+xml"/><item><title>Claude Desktop App Update: What's Worth Using</title><link>https://gabezen.com/blog/claude-desktop-app-update/</link><pubDate>Sun, 01 Nov 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-desktop-app-update/</guid><description>&lt;p&gt;The Claude desktop app changed a lot over the past two months, and nearly all of it lives in the Code tab. For a small company, four things deserve your time: sessions that can start and read other sessions, the context card, &lt;code&gt;/skill-doctor&lt;/code&gt;, and the shutdown of the old Claude Design site on December 14, which is the only item here with a deadline. Projects get the most attention, and they aren&amp;rsquo;t offered on Team or Enterprise plans yet.&lt;/p&gt;</description></item><item><title>Can I Put Client Data Into Claude?</title><link>https://gabezen.com/blog/can-i-put-client-data-into-claude/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/can-i-put-client-data-into-claude/</guid><description>&lt;p&gt;Yes, if the client&amp;rsquo;s contract allows it, you&amp;rsquo;re on Claude Team or Enterprise, and the data isn&amp;rsquo;t a kind that needs more than that. All three have to hold, and the training setting is the smallest part of it.&lt;/p&gt;&#10;&lt;p&gt;Anthropic&amp;rsquo;s own documents held two surprises. The word &amp;ldquo;confidential&amp;rdquo; appears nowhere in &lt;a href="https://www.anthropic.com/legal/consumer-terms"&gt;the Consumer Terms&lt;/a&gt;, so turning off training on Pro doesn&amp;rsquo;t get you a confidentiality promise. And on the current memory setup, deleting a chat doesn&amp;rsquo;t delete what Claude remembered from it.&lt;/p&gt;</description></item><item><title>Is Claude Code Safe to Use?</title><link>https://gabezen.com/blog/is-claude-code-safe/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-claude-code-safe/</guid><description>&lt;p&gt;Yes, if you treat it as a program running under your own login and you choose who approves its actions. The default approver may not be you. The name suggests a tool that edits code in one folder, but Anthropic&amp;rsquo;s docs say it can run &lt;a href="https://code.claude.com/docs/en/how-claude-code-works"&gt;&amp;ldquo;Any command you could run&amp;rdquo;&lt;/a&gt;.&lt;/p&gt;&#10;&lt;p&gt;Since late September, &lt;a href="https://code.claude.com/docs/en/permission-modes"&gt;new terminal and VS Code sessions start in auto mode&lt;/a&gt; on every plan. In auto mode, a second model reviews actions instead of you. In &lt;a href="https://www.anthropic.com/engineering/claude-code-auto-mode"&gt;Anthropic&amp;rsquo;s own evaluation&lt;/a&gt;, that model let through 17% of a small set of real overeager actions. And the folder limit people assume is, in &lt;a href="https://code.claude.com/docs/en/security"&gt;Anthropic&amp;rsquo;s words&lt;/a&gt;, &amp;ldquo;a permission prompt.&amp;rdquo;&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Dropbox?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-dropbox/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-dropbox/</guid><description>&lt;p&gt;Safe enough if you block &lt;code&gt;delete&lt;/code&gt; first, because the connector can remove your Dropbox files and can&amp;rsquo;t bring them back. &lt;a href="https://github.com/dropbox/dropbox-ai-plugins/blob/main/claude/README.md"&gt;Dropbox&amp;rsquo;s own notes for that server&lt;/a&gt; say &amp;ldquo;Revision history and restore workflows are not available until the required Dropbox MCP tools are exposed.&amp;rdquo; The undo belongs to a person at dropbox.com, and on a team the big undos belong to an admin.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 31, 2026.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Fireflies?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-fireflies/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-fireflies/</guid><description>&lt;p&gt;Connecting Claude to Fireflies is safe enough once you change one Fireflies default, and I wouldn&amp;rsquo;t do it before. By default every teammate in your workspace can open your meetings, including ones they weren&amp;rsquo;t in. Claude reaches what the person who connected it can reach.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything about Fireflies here comes from public pages I read on October 31, 2026.&lt;/p&gt;&#10;&lt;p&gt;The tool count depends on where you look. &lt;a href="https://claude.com/marketplace/connectors/fireflies"&gt;The listing in Claude&amp;rsquo;s directory&lt;/a&gt; shows three, all reads. &lt;a href="https://docs.fireflies.ai/mcp-tools/overview"&gt;Fireflies&amp;rsquo; reference&lt;/a&gt; for the same server address lists 19, including one that shares a meeting by email. &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Google Calendar?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-google-calendar/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-google-calendar/</guid><description>&lt;p&gt;Safe enough if every write tool stays on Needs approval and strangers&amp;rsquo; invites can&amp;rsquo;t land on your calendar by themselves. This connector also sends email. &lt;a href="https://developers.google.com/workspace/calendar/api/v3/reference/mcp/tools_list/update_event"&gt;By Google&amp;rsquo;s own reference&lt;/a&gt;, creating, changing, deleting or answering an event emails all attendees by default. &lt;a href="https://claude.com/marketplace/connectors/google-calendar"&gt;The listing in Claude&amp;rsquo;s directory&lt;/a&gt; never says so.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. This post comes from public pages I read on October 31, 2026. For company work we ruled it out. On April 2 we added a line to the instructions for our own Claude setup: &amp;ldquo;Never use MCP Gmail/Calendar/Drive connectors&amp;rdquo; (&lt;a href="https://gabezen.com/blog/claude-code-vs-cowork/"&gt;the full reasoning&lt;/a&gt;). We &lt;a href="https://gabezen.com/blog/claude-md-file-best-practices/"&gt;use a command-line tool for Calendar&lt;/a&gt; instead, because we wanted one login path and one tool to inspect when something goes wrong. That choice says nothing about whether the connector is safe, so there&amp;rsquo;s a test below.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Gusto?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-gusto/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-gusto/</guid><description>&lt;p&gt;Connecting Claude to Gusto is safe enough if one named admin connects it, approves every tool call and keeps other connectors out of that chat. &lt;a href="https://claude.com/marketplace/connectors/gusto"&gt;The listing&lt;/a&gt; undersells what you&amp;rsquo;re granting. It says &amp;ldquo;Query and analyze your Gusto data&amp;rdquo;. Its tools can also run payroll, link a bank account and create a company.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 31, 2026.&lt;/p&gt;&#10;&lt;p&gt;The listing has 62 tools. &lt;a href="https://docs.gusto.com/app-integrations/docs/mcp"&gt;Gusto&amp;rsquo;s docs&lt;/a&gt; describe 12 more that it doesn&amp;rsquo;t show, including IRS filings. &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Jira?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-jira/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-jira/</guid><description>&lt;p&gt;Safe enough if an Atlassian admin sets the permissions first. If nobody does, you&amp;rsquo;ve granted far more than the name suggests. You connect for Jira tickets, and the same connector reaches Confluence, Bitbucket, Loom and your service desk, with tools that make a page public and merge code.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 31, 2026.&lt;/p&gt;&#10;&lt;p&gt;&lt;a href="https://claude.com/marketplace/connectors/atlassian"&gt;The listing in Claude&amp;rsquo;s directory&lt;/a&gt; shows 21 tools. &lt;a href="https://support.atlassian.com/atlassian-ai-gateway/docs/supported-tools/"&gt;Atlassian&amp;rsquo;s docs&lt;/a&gt; name 100 more that can change something, and the full catalog runs to about 250. &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Notion?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-notion/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-notion/</guid><description>&lt;p&gt;Safe enough, if you first check what your Notion is connected to. On Business or Enterprise with Notion AI, your connection can search whatever your workspace has linked, such as Slack, Gmail, Google Drive and GitHub. Notion&amp;rsquo;s docs also list tools that trash whole databases and change calendar events, which may notify people.&lt;/p&gt;&#10;&lt;p&gt;Notion is our project tracker and document store. Our setup reaches it through API keys. I haven&amp;rsquo;t tested this connector. What I say about it comes from public pages I read on October 31, 2026.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Ramp?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-ramp/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-ramp/</guid><description>&lt;p&gt;Safe enough if you choose Read-only on Ramp&amp;rsquo;s sign-in screen. The default is Full access, which can approve bills, unlock cards and get card credentials to buy things. Claude&amp;rsquo;s directory calls this connector read-only. Ramp&amp;rsquo;s own guide describes a lot more.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 31, 2026.&lt;/p&gt;&#10;&lt;p&gt;&lt;a href="https://claude.com/marketplace/connectors/ramp"&gt;The listing in Claude&amp;rsquo;s directory&lt;/a&gt; says it &amp;ldquo;Enables secure, read-only access to Ramp data&amp;rdquo; and shows 17 tools. &lt;a href="https://docs.ramp.com/developer-api/v1/ramp-mcp"&gt;Ramp&amp;rsquo;s guide&lt;/a&gt; says the consent screen &amp;ldquo;asks you to choose &lt;strong&gt;Full access&lt;/strong&gt; (the default) or &lt;strong&gt;Read-only&lt;/strong&gt;.&amp;rdquo; &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Stripe?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-stripe/</link><pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-stripe/</guid><description>&lt;p&gt;Safe enough if you start in a sandbox and give the live account the fewest permissions Stripe&amp;rsquo;s sign-in screen allows. Stripe asks a person to confirm a refund. &lt;a href="https://docs.stripe.com/mcp"&gt;Its docs&lt;/a&gt; mark no such step for voiding an invoice or canceling a subscription through the same tool.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 31, 2026.&lt;/p&gt;&#10;&lt;p&gt;&lt;a href="https://claude.com/marketplace/connectors/stripe"&gt;The listing in Claude&amp;rsquo;s directory&lt;/a&gt; calls it &amp;ldquo;Payment processing and financial infrastructure tools&amp;rdquo; and never mentions writing. Its 11 tools include &lt;code&gt;stripe_api_write&lt;/code&gt;, which &lt;a href="https://docs.stripe.com/mcp"&gt;Stripe&amp;rsquo;s docs&lt;/a&gt; say can &amp;ldquo;Write data with any Stripe API &lt;code&gt;POST&lt;/code&gt;, &lt;code&gt;PATCH&lt;/code&gt;, &lt;code&gt;PUT&lt;/code&gt; and &lt;code&gt;DELETE&lt;/code&gt; method&amp;rdquo;. &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>AI Project Failure Rates: Where the 95% Comes From</title><link>https://gabezen.com/blog/ai-project-failure-rate/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-project-failure-rate/</guid><description>&lt;p&gt;Few of the famous &amp;ldquo;X% of AI projects fail&amp;rdquo; numbers measured AI projects failing. RAND&amp;rsquo;s 80% is RAND repeating &amp;ldquo;some estimates&amp;rdquo; from a 2022 magazine profile of an AI software CEO. MIT&amp;rsquo;s 95% is stated at least four different ways in a report that never shows the arithmetic, and the one chart it prints points nearer 75%. The IDC playbook behind the 88% never prints that number. It shows 4 launches out of 33 tries. Gartner&amp;rsquo;s 30% and 40% began as predictions. The studies with control groups and usage logs tell a duller story: AI speeds people up on work it suits, slows them down on work it doesn&amp;rsquo;t, and paid licences can sit unopened on any given day.&lt;/p&gt;</description></item><item><title>Can My Employer See My Claude Chats?</title><link>https://gabezen.com/blog/can-my-employer-see-my-claude-chats/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/can-my-employer-see-my-claude-chats/</guid><description>&lt;p&gt;Yes, if your seat is on your company&amp;rsquo;s Claude Team or Enterprise plan. Anthropic says nobody can browse your chats in the app, but at least one person can export the chats you haven&amp;rsquo;t deleted, incognito chats too. Keep private things off that seat. &lt;a href="https://support.claude.com/en/articles/9267276-roles-and-permissions"&gt;Settings, then Account&lt;/a&gt; shows your Primary Owner, who can export.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this or run an org export. It comes from Anthropic&amp;rsquo;s docs and a few vendor pages as of October 30, 2026, and it isn&amp;rsquo;t legal advice.&lt;/p&gt;</description></item><item><title>Is Claude Cowork Safe for Business Use?</title><link>https://gabezen.com/blog/is-claude-cowork-safe-for-business/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-claude-cowork-safe-for-business/</guid><description>&lt;p&gt;My call: yes for low-stakes office work (drafts, research, non-confidential documents), on a Team plan, in cloud sessions, from one backed-up folder, with the defaults below changed. For regulated work that needs an audit trail, my call is no on Team, because audit logs and the Compliance API are Enterprise only. Whether going without one is acceptable is for your lawyer. Keep financial files and anything under an NDA out of it.&lt;/p&gt;</description></item><item><title>Is Claude Team SOC 2 Compliant?</title><link>https://gabezen.com/blog/is-claude-team-soc-2-compliant/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-claude-team-soc-2-compliant/</guid><description>&lt;p&gt;Yes. &lt;a href="https://trust.anthropic.com/"&gt;Anthropic&amp;rsquo;s Trust Center&lt;/a&gt; lists Claude Team as in scope for SOC 2 Type 2. It marks HIPAA as N/A for Team, so that answer is no.&lt;/p&gt;&#10;&lt;p&gt;The catch is in the dates. The latest report Anthropic publishes &lt;a href="https://trust.anthropic.com/doc/trust?rid=6925de671c0c91e594eed26b&amp;amp;r=iz673w96495gyjer8h78n"&gt;covers October 1, 2024 to September 30, 2025&lt;/a&gt;. &lt;a href="https://support.claude.com/en/articles/12138966-release-notes"&gt;Cowork&amp;rsquo;s first release came on January 12, 2026&lt;/a&gt;. So a Claude Team seat today includes a feature that didn&amp;rsquo;t exist during the audited period. Write &amp;ldquo;all of it is SOC 2 audited&amp;rdquo; on a customer&amp;rsquo;s questionnaire and you&amp;rsquo;re claiming more than the published report shows.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Google Drive?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-google-drive/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-google-drive/</guid><description>&lt;p&gt;Probably, if you clean up old sharing, block the share_file and trash_file tools, and keep Claude off files from strangers. Claude&amp;rsquo;s directory lists tools to share and trash files, and you should assume it can read nearly anything your account can open.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. This post is what I read on October 30: Anthropic&amp;rsquo;s and Google&amp;rsquo;s pages and the security research linked below. For company work we went another way. On April 2 we added a line to the standing instructions for our own Claude setup: &amp;ldquo;Never use MCP Gmail/Calendar/Drive connectors&amp;rdquo; (&lt;a href="https://gabezen.com/blog/claude-code-vs-cowork/"&gt;the full reasoning&lt;/a&gt;). We wanted one login path and one tool to inspect when something goes wrong. That says nothing about whether the connector is safe, so this post includes a test.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to HubSpot?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-hubspot/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-hubspot/</guid><description>&lt;p&gt;Safe enough for CRM work, if you set the write tools to Needs approval before anyone uses it and decide who owns anything public, because it can also put things on your website.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector. Everything here comes from public pages I read on October 30, 2026.&lt;/p&gt;&#10;&lt;p&gt;&lt;a href="https://claude.com/marketplace/connectors/hubspot"&gt;Claude&amp;rsquo;s directory&lt;/a&gt; lists 36 tools and sums the connector up as &amp;ldquo;Search and update contacts, companies, deals, tickets, campaigns, and more&amp;rdquo;. That blurb leaves out publishing. &lt;a href="https://knowledge.hubspot.com/integrations/set-up-and-use-the-hubspot-connector-for-claude"&gt;HubSpot&amp;rsquo;s own help page&lt;/a&gt; gives &amp;ldquo;draft and publish a blog post&amp;rdquo; as its example prompt. A sales team connecting this for pipeline reviews also connects the company blog, if the person connecting has the permissions. &lt;a href="https://gabezen.com/connectors/"&gt;The table of what 33 connectors can change&lt;/a&gt; compares it with your other apps.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Microsoft 365?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-microsoft-365/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-microsoft-365/</guid><description>&lt;p&gt;I&amp;rsquo;d connect Claude to Microsoft 365 for reading, once you&amp;rsquo;ve checked write tools are off and searched SharePoint as your most junior employee. We run on Google and I haven&amp;rsquo;t tested this connector, so this comes from the docs as of October 30, 2026.&lt;/p&gt;&#10;&lt;h2 id="is-the-claude-microsoft-365-connector-read-only"&gt;Is the Claude Microsoft 365 connector read-only?&lt;/h2&gt;&#10;&lt;p&gt;&lt;a href="https://support.claude.com/en/articles/12542951-set-up-the-microsoft-365-connector"&gt;Read-only by default, Anthropic says&lt;/a&gt;, but go and look. &lt;a href="https://claude.com/marketplace/connectors/microsoft-365"&gt;The directory listing&lt;/a&gt; shows seven tools, all reads. &lt;a href="https://support.claude.com/en/articles/12684923-microsoft-365-connector-security-guide"&gt;Anthropic&amp;rsquo;s security guide&lt;/a&gt; lists 35 write tools on top of those, available once an organization enables them.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Shopify?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-shopify/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-shopify/</guid><description>&lt;p&gt;Safe enough for a working store if you connect through a staff login limited to what Claude should touch, and try the write tools on a development store first. Once approved, the connector can change your products, prices and stock and create discount codes. Shopify says it can&amp;rsquo;t refund, cancel an order or change your plan, payments or taxes.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested this connector on a store. Everything here comes from public pages and the connector&amp;rsquo;s public metadata, read on October 30, 2026.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Slack?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-slack/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-slack/</guid><description>&lt;p&gt;Safe enough if an owner sets its tools first, and riskier than it looks if you stop at blocking private search. Slack&amp;rsquo;s channel and thread tools read private messages too, and the same connector can send messages on your behalf.&lt;/p&gt;&#10;&lt;p&gt;We use Slack at my company, and &lt;a href="https://gabezen.com/blog/claude-code-for-project-managers/"&gt;one of our Claude Code commands used to DM engineers for approval&lt;/a&gt; until we cut that step. I haven&amp;rsquo;t tested this connector. Everything about it here comes from public pages I read on October 30, 2026.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Xero?</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-xero/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-xero/</guid><description>&lt;p&gt;Reasonably safe, going by what Xero publishes, if you use its read-only connector and stay away from its two MCP routes. Read the consent screen before you click Allow. After that, put dates in every question and check any number you&amp;rsquo;d act on against Xero.&lt;/p&gt;&#10;&lt;p&gt;I haven&amp;rsquo;t tested any of this. My company has never used Xero. We run on QuickBooks, which &lt;a href="https://gabezen.com/blog/ai-bookkeeping-small-business/"&gt;we moved to in August&lt;/a&gt;. Everything here comes from public pages I read on October 30, 2026.&lt;/p&gt;</description></item><item><title>Is It Safe to Upload Bank Statements to Claude?</title><link>https://gabezen.com/blog/is-it-safe-to-upload-bank-statements-to-claude/</link><pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-upload-bank-statements-to-claude/</guid><description>&lt;p&gt;It&amp;rsquo;s safe enough for your own statements if you strip the account numbers, keep the chat out of training and memory, and check every total against the statement yourself. Someone else&amp;rsquo;s statements are a different question, and on a personal plan I&amp;rsquo;d say no.&lt;/p&gt;&#10;&lt;p&gt;We ran 15 months of bank statement PDFs through Claude Code before we moved to QuickBooks, and the accuracy section comes from that. Claude Code skips the chat upload button, so I didn&amp;rsquo;t test the chat settings below. They come from public pages I read on October 30, 2026.&lt;/p&gt;</description></item><item><title>How to Set Up Claude's Free Small Business Plugin</title><link>https://gabezen.com/blog/claude-small-business-plugin/</link><pubDate>Thu, 29 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-small-business-plugin/</guid><description>&lt;p&gt;If you pay for Claude and it still doesn&amp;rsquo;t know your business, start here. Anthropic gives away a plugin called Claude for Small Business. It comes with 44 skills for proposals, follow-ups, invoices, hiring and marketing, and a setup that interviews you about your business so the skills stop guessing. It costs nothing on top of a Pro, Max, Team or Enterprise plan, and the setup takes about 15 minutes.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect ChatGPT to Gmail? What to Check</title><link>https://gabezen.com/blog/is-it-safe-to-connect-chatgpt-to-gmail/</link><pubDate>Wed, 28 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-chatgpt-to-gmail/</guid><description>&lt;p&gt;Mostly yes, as long as you know what you&amp;rsquo;re granting before you click Allow. The bigger risk is what it can do after it reads, and who in your company gets to decide that.&lt;/p&gt;&#10;&lt;p&gt;A disclosure first: I run my company&amp;rsquo;s email through Claude, not ChatGPT, so this isn&amp;rsquo;t a review from daily use. It&amp;rsquo;s what ChatGPT&amp;rsquo;s Gmail connection can do according to OpenAI&amp;rsquo;s help pages and the guides that tested it, where they disagree, and how to settle it yourself in two minutes, because Google tells you exactly what any app can touch.&lt;/p&gt;</description></item><item><title>Claude Code as a Chief of Staff: 39 Runs, and Where Claude Duet Wins</title><link>https://gabezen.com/blog/claude-code-chief-of-staff/</link><pubDate>Sun, 25 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-chief-of-staff/</guid><description>&lt;p&gt;You can make Claude Code your chief of staff, and we did. Ours is a command each founder runs at the end of the day. It reads the day&amp;rsquo;s meetings and email, files what matters into the right client folder, finds replies we owe, proposes tasks and lines up tomorrow. Between June 4 and September 30 it ran 39 times across my co-founder and me, read 460 meetings and 1,420 email threads, and surfaced 626 threads where someone outside the company was waiting on us. The short version of how to build one is &lt;a href="#how-to-make-claude-code-your-chief-of-staff"&gt;at the end&lt;/a&gt;: a folder, one command and two weeks of running it by hand.&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to QuickBooks? What It Can Touch</title><link>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-quickbooks/</link><pubDate>Sun, 25 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/is-it-safe-to-connect-claude-to-quickbooks/</guid><description>&lt;p&gt;Connecting Claude to QuickBooks is safe enough if you decide what it can touch before you click Allow. Most people don&amp;rsquo;t, because the screen says &amp;ldquo;QuickBooks&amp;rdquo; and they picture someone reading their P&amp;amp;L. Intuit&amp;rsquo;s official connector does much more than read. Since July it can create, update, send and delete invoices and estimates, set up payment links and add customers, according to &lt;a href="https://www.intuit.com/blog/news-social/quickbooks-expands-into-claude-and-chatgpt-with-new-sales-invoicing-payroll-and-lending-features/"&gt;Intuit&amp;rsquo;s own announcement&lt;/a&gt;. Its listing in Claude&amp;rsquo;s directory adds tools that import transactions into your books and update your company profile. Reading reports is the small part of the grant.&lt;/p&gt;</description></item><item><title>How to Train ChatGPT on Your Own Data, and When Not To</title><link>https://gabezen.com/blog/train-chatgpt-on-company-data/</link><pubDate>Tue, 20 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/train-chatgpt-on-company-data/</guid><description>&lt;p&gt;&amp;ldquo;Can I train ChatGPT on our company data?&amp;rdquo; is the most common way people ask for an AI that knows their business. It&amp;rsquo;s also slightly the wrong question. Training changes the model itself. What almost everyone wants is a model that looks up the right document before it answers, and tells you which document it used. That second thing is cheaper, easier to keep current, and it&amp;rsquo;s what every working setup I know of actually does.&lt;/p&gt;</description></item><item><title>How Much Does an AI Chatbot for Your Business Cost? 56 Gigs</title><link>https://gabezen.com/blog/hire-someone-ai-knowledge-base/</link><pubDate>Sun, 18 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/hire-someone-ai-knowledge-base/</guid><description>&lt;p&gt;If you search for someone to set up an AI that knows your company, you&amp;rsquo;ll find prices from $30 to $20,000 for what sounds like the same thing. Most of the difference is about what happens after the first demo: whether the answers stay current, whether you can check where they came from, and whether you still own the thing when the freelancer moves on.&lt;/p&gt;&#10;&lt;p&gt;For the internal kind, the one your team asks about clients and past decisions, the short answer from what I found: the $500 to $3,000 gigs usually load one batch of your documents once, and the ones that connect email, calls and shared drives on a schedule start around $4,000 and run to $20,000. What moves the price is how many sources it connects, how many people need different access, and who keeps it running after delivery.&lt;/p&gt;</description></item><item><title>Claude Code vs Cowork: Why Our Team Still Uses the Terminal</title><link>https://gabezen.com/blog/claude-code-vs-cowork/</link><pubDate>Sat, 17 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-vs-cowork/</guid><description>&lt;p&gt;Claude Code and Cowork run the same agent. Cowork lives in the Claude app: you give it a folder or a connector, describe the result you want, and it hands back finished files, with no terminal involved. Claude Code works inside a folder you start it in, usually a git repository, where its instructions, commands and safety checks are ordinary files your whole team can review and share. As of October 2026, both come with every paid Claude plan.&lt;/p&gt;</description></item><item><title>Custom GPTs Are Retiring: What to Use for Your Business</title><link>https://gabezen.com/blog/custom-gpt-for-business/</link><pubDate>Sat, 17 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/custom-gpt-for-business/</guid><description>&lt;p&gt;OpenAI is retiring custom GPTs on December 11, 2026, and moving them to ChatGPT plugins (&lt;a href="https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq"&gt;its retirement and migration FAQ&lt;/a&gt;). If your business leans on a GPT, you have until then to migrate it to a plugin, rebuild it as a Claude Project, or replace it with something a team can share.&lt;/p&gt;&#10;&lt;p&gt;Which one depends on who uses it. A custom GPT was a fine choice if one person used it and it answered from a handful of documents that rarely change, like a rate card, a handbook or a style guide. The plugin version will be fine for that too. It starts failing once a whole team depends on it. The files are a snapshot from the day someone uploaded them. Answers don&amp;rsquo;t point back to a source unless you write that into the instructions. And everyone the GPT is shared with reads the same files, so you can&amp;rsquo;t give the sales lead and the new contractor different views. Migrating keeps all three problems.&lt;/p&gt;</description></item><item><title>Claude Code for Non-Developers: Four Installs, One Team</title><link>https://gabezen.com/blog/claude-code-for-non-developers/</link><pubDate>Fri, 16 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-for-non-developers/</guid><description>&lt;p&gt;Claude Code is good for non-developers when most of their work lives in documents: client notes, plans, finance trackers, weekly updates. &lt;a href="https://code.claude.com/docs/en/overview"&gt;Claude Code&lt;/a&gt; reads and edits files in a folder, and anyone who can describe what they want can direct it. It gets harder when you need it to stay out of the terminal completely, or when nobody on the team can review the parts that behave like software.&lt;/p&gt;&#10;&lt;p&gt;Our team has very different technical backgrounds, and all of us work out of one repository that I call our &lt;a href="https://gabezen.com/blog/company-brain-that-runs-a-company/"&gt;company brain&lt;/a&gt; (&lt;a href="https://gabezen.com/blog/company-brain-claude-code-setup/"&gt;how that repo is laid out&lt;/a&gt;).&lt;/p&gt;</description></item><item><title>Is It Safe to Connect Claude to Gmail? What It Can Touch</title><link>https://gabezen.com/blog/can-claude-read-my-email/</link><pubDate>Thu, 15 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/can-claude-read-my-email/</guid><description>&lt;p&gt;Yes, if you check four things first: what it&amp;rsquo;s allowed to do after it reads, who has to approve that, who on your team can change the approval, and what happens to the mail it pulls in. Claude&amp;rsquo;s Gmail connector has a clear answer to all four, and one of them is a setting your company owner controls, not you.&lt;/p&gt;&#10;&lt;p&gt;I&amp;rsquo;ve been on both sides of this. Our company read its email with Claude through a system we built ourselves, starting in April, and it has filed more than 1,200 of our threads. The official Gmail connector now does most of what we built, so this post covers what the connector can touch, what we learned running our own, and the checklist I&amp;rsquo;d use on any AI that wants your inbox.&lt;/p&gt;</description></item><item><title>A Company Brain Example: The One That Runs My Company</title><link>https://gabezen.com/blog/company-brain-that-runs-a-company/</link><pubDate>Wed, 14 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/company-brain-that-runs-a-company/</guid><description>&lt;p&gt;A company brain is one place where AI can read your company&amp;rsquo;s email, call transcripts and documents, and answer questions with a link to the source (&lt;a href="https://gabezen.com/company-brain/"&gt;the longer definition, with its parts&lt;/a&gt;).&lt;/p&gt;&#10;&lt;p&gt;My company runs on one git repository. It holds every client folder, every call transcript we&amp;rsquo;re allowed to keep, a few thousand email syntheses, and a set of Claude Code commands that read all of it. When my co-founder asks &amp;ldquo;what did we promise this client about hosting?&amp;rdquo;, the answer comes back with a link to the email or the transcript line where we said it. That repository is our company brain.&lt;/p&gt;</description></item><item><title>AI Note Taker Privacy: 46 Transcripts Got Past Our Gate</title><link>https://gabezen.com/blog/confidential-transcripts-reached-main/</link><pubDate>Tue, 13 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/confidential-transcripts-reached-main/</guid><description>&lt;p&gt;Two days after we shipped a four-layer confidentiality gate for meeting transcripts, a documentation pull request merged 46 transcripts that gate would have held. The gate itself worked fine. It just wasn&amp;rsquo;t standing anywhere the pull request had to pass.&lt;/p&gt;&#10;&lt;p&gt;If you&amp;rsquo;re deciding whether an AI note taker is safe for confidential meetings, the three rules I&amp;rsquo;d give you are &lt;a href="#what-id-tell-you-before-you-connect-your-meetings"&gt;at the end&lt;/a&gt;. The rest is how we learned them.&lt;/p&gt;&#10;&lt;p&gt;I&amp;rsquo;ll describe what was in those files in categories only, for reasons that will be obvious.&lt;/p&gt;</description></item><item><title>Self-Improving AI: Our 3 Feedback Loops Collected Nothing</title><link>https://gabezen.com/blog/feedback-loops-collected-nothing/</link><pubDate>Tue, 13 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/feedback-loops-collected-nothing/</guid><description>&lt;p&gt;In August we built three ways for our company brain to learn from its own work: an estimate calibration engine, a log of which skills get used, and a hook that asks &amp;ldquo;is this worth building?&amp;rdquo; before we build something. Two months later the calibration store has zero rows, the usage log has zero lines, and the hook never fired once on the day it went live. Meanwhile, the one loop our planning doc described as fully closed turned out never to have run at all.&lt;/p&gt;</description></item><item><title>How to Build a Company Brain in Claude Code (and What Broke)</title><link>https://gabezen.com/blog/company-brain-claude-code-setup/</link><pubDate>Mon, 12 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/company-brain-claude-code-setup/</guid><description>&lt;p&gt;A company brain in Claude Code (&lt;a href="https://gabezen.com/company-brain/"&gt;what a company brain is&lt;/a&gt;) starts as a folder of markdown files with a &lt;code&gt;CLAUDE.md&lt;/code&gt; at the root. The &lt;code&gt;CLAUDE.md&lt;/code&gt; tells Claude what&amp;rsquo;s in the folder, where new things go and which rules to follow. Add a few slash commands for the jobs you repeat, like turning a meeting transcript into notes, and you have the first version. If your notes already live in an Obsidian vault, the vault can be the folder (&lt;a href="https://obsidian.md/help/data-storage"&gt;Obsidian&amp;rsquo;s own docs&lt;/a&gt; describe a vault as a folder of plain-text markdown files): open Claude Code inside it and write the &lt;code&gt;CLAUDE.md&lt;/code&gt;.&lt;/p&gt;</description></item><item><title>CLAUDE.md Best Practices: What Our 223-Line File Taught Us</title><link>https://gabezen.com/blog/claude-md-file-best-practices/</link><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-md-file-best-practices/</guid><description>&lt;p&gt;A &lt;code&gt;CLAUDE.md&lt;/code&gt; file is the instructions file Claude Code reads at the start of every session. The short version of what works for us:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;&lt;strong&gt;Put in&lt;/strong&gt; what the folder is for, a map of where things go, and the &amp;ldquo;always&amp;rdquo; and &amp;ldquo;never&amp;rdquo; rules you&amp;rsquo;d otherwise repeat in chat.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Leave out&lt;/strong&gt; step-by-step procedures (those are skills or commands), lists of your tools, anything that changes weekly, and secrets.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Location:&lt;/strong&gt; one at the root of the project folder, one in each department folder, and your personal preferences in &lt;code&gt;~/.claude/CLAUDE.md&lt;/code&gt;.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Size:&lt;/strong&gt; Anthropic&amp;rsquo;s docs say to &lt;a href="https://code.claude.com/docs/en/memory"&gt;target under 200 lines per file&lt;/a&gt;. Ours is 223, and the extra lines cost us.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;When to add a line:&lt;/strong&gt; the second time Claude makes the same mistake.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;Everything below is the evidence. Our whole company runs out of one git repository (the &lt;a href="https://gabezen.com/blog/company-brain-that-runs-a-company/"&gt;company brain&lt;/a&gt;), and its root &lt;code&gt;CLAUDE.md&lt;/code&gt; has been edited 79 times since March. Most of the people who use it aren&amp;rsquo;t developers (&lt;a href="https://gabezen.com/blog/claude-code-for-non-developers/"&gt;how that works for them&lt;/a&gt;), so the file has to do the explaining a senior teammate would otherwise do.&lt;/p&gt;</description></item><item><title>Decision Log: Write Down the Rulings Your AI Follows</title><link>https://gabezen.com/blog/write-down-the-rulings/</link><pubDate>Sat, 10 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/write-down-the-rulings/</guid><description>&lt;p&gt;When we started logging our company&amp;rsquo;s rulings in July, six of the first ones were copied out of one person&amp;rsquo;s private AI memory files. One of them was the rule that the AI never sends client email on its own. It had been quietly governing that person&amp;rsquo;s Claude sessions and nobody else&amp;rsquo;s. The log we built to fix that is the part of our &lt;a href="https://gabezen.com/blog/company-brain-that-runs-a-company/"&gt;company brain&lt;/a&gt; I&amp;rsquo;d build first if I were starting over.&lt;/p&gt;</description></item><item><title>What Is Tribal Knowledge? Five Places Ours Hid From Our AI</title><link>https://gabezen.com/blog/tribal-knowledge-ai/</link><pubDate>Fri, 09 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/tribal-knowledge-ai/</guid><description>&lt;p&gt;Tribal knowledge is what people in a company know but never wrote down: how a process actually runs, why a price is what it is, who has the login to the payment processor. Institutional knowledge is the bigger set, everything the company knows, including what&amp;rsquo;s written in SOPs, contracts and wikis. Tribal knowledge is the slice that leaves when a person leaves. Some people avoid the term because of its baggage and say &amp;ldquo;undocumented knowledge&amp;rdquo; or &amp;ldquo;institutional memory&amp;rdquo; instead, which is fine. The problem is the same under any name. Adding AI changes the stakes in one specific way. An AI can only answer from what it can read, and when the written copy is wrong, it answers wrong with complete confidence.&lt;/p&gt;</description></item><item><title>We Deleted 52 Claude Code Skills in Three Months</title><link>https://gabezen.com/blog/deleted-52-skills/</link><pubDate>Thu, 08 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/deleted-52-skills/</guid><description>&lt;p&gt;AI made it nearly free to add tooling to our &lt;a href="https://gabezen.com/blog/company-brain-that-runs-a-company/"&gt;company brain&lt;/a&gt;, so in March 2026 we added it fast. Between March and May we deleted 52 commands, skills and agents. The one we actually missed came back in June, rewritten for a real project instead of an imagined one, and it&amp;rsquo;s better for having been deleted. If your own &lt;code&gt;.claude/&lt;/code&gt; folder is crowded, the rules we use now are under &lt;a href="#how-we-decide-what-to-delete"&gt;how we decide what to delete&lt;/a&gt;.&lt;/p&gt;</description></item><item><title>AI for Onboarding New Employees: What Worked on Three Hires</title><link>https://gabezen.com/blog/ai-employee-onboarding/</link><pubDate>Wed, 07 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-employee-onboarding/</guid><description>&lt;p&gt;AI is good at writing an onboarding kit from documents you already have, and at noticing when two of those documents disagree. It&amp;rsquo;s bad at knowing what actually happened during setup. We learned both between late August and late September, when we brought on three people: a part-time bookkeeper, a fractional CFO and an outbound sales lead.&lt;/p&gt;&#10;&lt;p&gt;Onboarding is mostly a knowledge problem. A new person needs to know who owns what, how things are done here, which documents to trust, and what to do first. In most small companies that knowledge comes from whichever founder has time that week, and it shows. Gallup has reported that &lt;a href="https://www.gallup.com/workplace/235121/why-onboarding-experience-key-retention.aspx"&gt;only 12% of employees strongly agree&lt;/a&gt; their organization does a great job of onboarding new employees.&lt;/p&gt;</description></item><item><title>Why AI Agents Fail in Production: 8 Times Our Tools Lied</title><link>https://gabezen.com/blog/small-ways-tools-lied/</link><pubDate>Tue, 06 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/small-ways-tools-lied/</guid><description>&lt;p&gt;Most of the AI agent failures we&amp;rsquo;ve hit in production started in the tools feeding our company brain. The model itself caused far fewer. Transcription, email, meeting notes, the task database and authentication each returned a confident answer that was wrong, stale or empty, and none of them raised an error. Here are eight from the last few months and the check we added for each one.&lt;/p&gt;&#10;&lt;p&gt;Some context. Every evening, a pipeline reads the day&amp;rsquo;s meetings and email for each of the two founders, files what matters into the right client folder, and proposes tasks. Between June and September it processed 460 meetings and 1,420 email threads. At that volume, a tool that is wrong 1% of the time is wrong several times a week. The model downstream has no way to know, because the input looks normal.&lt;/p&gt;</description></item><item><title>Can Claude Send Emails for You? Ours Drafts and Never Sends</title><link>https://gabezen.com/blog/ai-drafts-every-email-sends-none/</link><pubDate>Mon, 05 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-drafts-every-email-sends-none/</guid><description>&lt;p&gt;Yes. Since August 2026, Claude&amp;rsquo;s Gmail connector can send, reply and forward, and it asks before each one by default. Our company brain could send too. Over four months it surfaced 626 threads where we owed a reply, it drafts answers on request, and a person presses send on every one. Every client email it writes lands in Gmail as a draft, and nothing goes out on the AI&amp;rsquo;s say-so.&lt;/p&gt;</description></item><item><title>AI Bookkeeping With Claude: 15 Months, Then QuickBooks</title><link>https://gabezen.com/blog/ai-bookkeeping-small-business/</link><pubDate>Sun, 04 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-bookkeeping-small-business/</guid><description>&lt;p&gt;AI can do the categorizing part of bookkeeping well. We built our own bookkeeper on Claude Code, and it sorted 559 bank transactions across 15 months of statements into our chart of accounts with zero left uncategorized. In August we retired it anyway and moved to QuickBooks Online Plus with a live bank feed. The hard part of keeping books turned out to be getting transactions in every day and keeping one set of books a teammate can close each month. A product we were about to pay for already did both.&lt;/p&gt;</description></item><item><title>AI for Writing SOPs: Draft From What Actually Happened</title><link>https://gabezen.com/blog/ai-for-sops/</link><pubDate>Sat, 03 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-for-sops/</guid><description>&lt;p&gt;Ask an AI to &amp;ldquo;write an SOP for client onboarding&amp;rdquo; and you&amp;rsquo;ll get a clean, generic document with a purpose section, a scope section and ten sensible steps that could belong to any company on earth. Nobody will follow it. It says nothing about how onboarding works at your company, or where it went wrong last time.&lt;/p&gt;&#10;&lt;p&gt;The SOPs that actually changed how we work came from the opposite direction. The AI read what happened on real projects (call transcripts, retrospectives, emails) and drafted a procedure from that. Then a person made the decisions the draft couldn&amp;rsquo;t make.&lt;/p&gt;</description></item><item><title>Claude Code Use Cases for Business: What Our 66 Commands Do</title><link>https://gabezen.com/blog/claude-code-use-cases-for-business/</link><pubDate>Fri, 02 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-use-cases-for-business/</guid><description>&lt;p&gt;The Claude Code use cases that pay off for a business are the ones where documents go in and documents come out. These are the ones that earn their keep at our company:&lt;/p&gt;&#10;&lt;table&gt;&#10;&#9;&lt;thead&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;Use case&lt;/th&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;What goes in, and what comes out&lt;/th&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/thead&gt;&#10;&#9;&lt;tbody&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Proposal&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Call notes, solution doc and estimate in. A 4 to 6 page proposal out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Weekly client update&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Each engineer&amp;rsquo;s rough notes in. One email per client out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;End of day&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Meetings and email in. Tasks, filed notes and a plan for tomorrow out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Cash reconciliation&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Bank ledger plus invoice and payment records in. A report and proposed fixes out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;LinkedIn campaign&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Your copy and a lead list in. A draft campaign, then drafted replies, out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Lead follow-up&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;A warm lead list in. Ranked Gmail drafts out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Timesheet&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Time tracker data and calendar in. Time entries for approval out.&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/tbody&gt;&#10;&lt;/table&gt;&#10;&lt;p&gt;None of them ask the person running them to write code.&lt;/p&gt;</description></item><item><title>AI Meeting Action Items: Ours Made 1,042 Tasks in 4 Months</title><link>https://gabezen.com/blog/ai-created-1042-tasks/</link><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-created-1042-tasks/</guid><description>&lt;p&gt;Between June 4 and September 30, our company brain created 1,042 tasks in our project tracker. It read our email and meeting transcripts at the end of each day, pulled out every meeting action item, and filed each one as its own task. Most of those tasks were noise, and the fix that finally worked was taking the job of counting away from the model entirely.&lt;/p&gt;&#10;&lt;h2 id="what-the-day-end-run-does"&gt;What the day-end run does&lt;/h2&gt;&#10;&lt;p&gt;My co-founder and I each run a command at the end of the day (we call it day-end, and I wrote up &lt;a href="https://gabezen.com/blog/claude-code-chief-of-staff/"&gt;how it works as a chief of staff&lt;/a&gt;). It reads the day&amp;rsquo;s meetings and email threads, figures out who owes what to whom, updates project docs, syncs contacts, and drafts tasks. A gate shows us the drafts before anything gets written, and we approve, edit or drop them.&lt;/p&gt;</description></item><item><title>Two Founders, One Company Brain, and the Duplicate Problem</title><link>https://gabezen.com/blog/two-founders-one-brain/</link><pubDate>Wed, 30 Sep 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/two-founders-one-brain/</guid><description>&lt;p&gt;My co-founder and I each run an AI end-of-day process on our own inbox and calendar, and most of our client meetings and threads include both of us. Left alone, two runs means every transcript gets filed twice and every action item becomes two tasks. Getting two runs to share one brain took a shared table, two kinds of keys, and one rule about owed replies that turned out to matter more than the rest.&lt;/p&gt;</description></item><item><title>Claude Code for Project Managers: The Weekly Update We Kept</title><link>https://gabezen.com/blog/claude-code-for-project-managers/</link><pubDate>Sun, 27 Sep 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-for-project-managers/</guid><description>&lt;p&gt;A project manager can hand Claude Code the assembly work: turning engineers&amp;rsquo; notes into a weekly client update, checking reported hours against the time tracker, drafting a retrospective from the project&amp;rsquo;s documents, and logging milestones once an update goes out. What stays human is judgment about the client (what to say, what to leave out, when something is going badly) and the send button. In our company, every client email Claude writes lands as a Gmail draft, and a person sends it.&lt;/p&gt;</description></item><item><title>Should I Automate This? First Ask Whether It Should Exist</title><link>https://gabezen.com/blog/should-it-exist-before-automating/</link><pubDate>Fri, 25 Sep 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/should-it-exist-before-automating/</guid><description>&lt;p&gt;Before you automate something, ask what happens if nobody does it at all. If the answer is nothing, delete it instead.&lt;/p&gt;&#10;&lt;p&gt;Our own build gate skipped that question at first. Its first verdict was DON&amp;rsquo;T, for a small job that would remind the team about upcoming vacations. The arithmetic said it would lose about $6,200 a year. The verdict was right. The reasoning skipped a question, and I patched the gate the same night.&lt;/p&gt;</description></item><item><title>Two Bots Argued on a Pull Request for 21 Hours</title><link>https://gabezen.com/blog/two-bots-argued-21-hours/</link><pubDate>Tue, 22 Sep 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/two-bots-argued-21-hours/</guid><description>&lt;p&gt;On the night of July 31, one of our scheduled AI routines opened a pull request, and our AI code reviewer left comments on it. The routine fixed what the reviewer flagged and pushed. The push triggered a new review. The new review found something else. They kept going for about 21 hours, through 147 review runs, and the bill showed up as roughly $500 in automatic credit reloads on our API account. Neither bot was broken. Each one did its job, and that&amp;rsquo;s why it took a day to stop.&lt;/p&gt;</description></item><item><title>Claude Code Team Setup: Shared Settings, /setup, What Broke</title><link>https://gabezen.com/blog/claude-code-team-setup/</link><pubDate>Sat, 19 Sep 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/claude-code-team-setup/</guid><description>&lt;p&gt;To set up Claude Code for a team that isn&amp;rsquo;t technical, put everything shared in one git repository: the shared settings file, the commands and the &lt;a href="https://gabezen.com/blog/claude-md-file-best-practices/"&gt;&lt;code&gt;CLAUDE.md&lt;/code&gt; instructions&lt;/a&gt;. Then write a setup command that installs what&amp;rsquo;s missing and asks each person only for the keys their role needs. Secrets go in a gitignored &lt;code&gt;.env&lt;/code&gt;. Plan for whatever login expires every day. And decide up front whether your Windows users run WSL or native Windows, because the install is different for each and the person who has never opened a terminal will hit the difference first.&lt;/p&gt;</description></item><item><title>Knowledge Graph vs RAG: We Piloted a Graph and Killed It</title><link>https://gabezen.com/blog/knowledge-graph-tested-and-killed/</link><pubDate>Thu, 06 Aug 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/knowledge-graph-tested-and-killed/</guid><description>&lt;p&gt;We gave a knowledge graph tool a fair pilot on 46 real documents, with the kill criteria written down before the first run. It drew mostly true edges. Then we ran it a second time on identical input and only 15.85% of the node labels matched. We turned it off, kept a boring search index, and I think that was the right call.&lt;/p&gt;&#10;&lt;h2 id="the-problem-we-were-actually-trying-to-solve"&gt;The problem we were actually trying to solve&lt;/h2&gt;&#10;&lt;p&gt;Our company brain is a git repo full of markdown. In July 2026 it held about 3,548 markdown files, including 395 meeting transcripts and 545 email threads. Retrieval was grep and reading files. That works when you know the word you&amp;rsquo;re looking for, and fails when you don&amp;rsquo;t.&lt;/p&gt;</description></item><item><title>How to Prompt Claude Code: 5 Lessons From 356 Sessions</title><link>https://gabezen.com/blog/making-of-guide/</link><pubDate>Mon, 09 Feb 2026 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/making-of-guide/</guid><description>&lt;p&gt;I was trying to build a procedural music engine in Web Audio API. Everything Claude gave me sounded like a ringtone. I kept prompting &amp;ldquo;make it better,&amp;rdquo; &amp;ldquo;add more depth,&amp;rdquo; &amp;ldquo;make it more emotional.&amp;rdquo; Three prompts that produce absolutely nothing, every time.&lt;/p&gt;&#10;&lt;p&gt;Then I wrote this: &amp;ldquo;Think about the upgrade from NES to SNES. Less gameboy, more super famicom. Think donkey kong country samples, the depth of a JRPG soundtrack, the epic surreal beauty of a PC-98 game.&amp;rdquo;&lt;/p&gt;</description></item><item><title>Microservices vs Monolith: What Segment's 140 Services Cost</title><link>https://gabezen.com/blog/microservices-coordination-tax/</link><pubDate>Sun, 28 Dec 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/microservices-coordination-tax/</guid><description>&lt;p&gt;Twilio Segment did everything &amp;ldquo;right.&amp;rdquo;&lt;/p&gt;&#10;&lt;p&gt;Every destination they sent data to got its own service, its own queue and eventually its own repo, with shared libraries holding it all together.&lt;/p&gt;&#10;&lt;p&gt;On paper it looked clean.&#10;In practice it was coordination hell.&lt;/p&gt;&#10;&lt;p&gt;Three full-time engineers spent most of their time just keeping the system alive. In Segment&amp;rsquo;s words, &amp;ldquo;as our velocity plummeted, our defect rate exploded,&amp;rdquo; and on-call engineers kept getting paged for load spikes.&lt;/p&gt;</description></item><item><title>Software Craftsmanship in the Age of AI: When Craft Pays</title><link>https://gabezen.com/blog/ai-craft-vs-outcomes/</link><pubDate>Sun, 28 Dec 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-craft-vs-outcomes/</guid><description>&lt;p&gt;There&amp;rsquo;s a divide in software development that&amp;rsquo;s always existed but rarely gets named. Some developers fall in love with the craft itself, the elegance of a well-factored system, the satisfaction of a clean abstraction. Others just want to ship something that works and move on to the next problem.&lt;/p&gt;&#10;&lt;p&gt;The sharpest example I&amp;rsquo;ve seen was a nightly CSV import. One developer spent most of a day on a beautifully architected pipeline, another had a working AI-generated version in 20 minutes, and the 20-minute version was the right call.&lt;/p&gt;</description></item><item><title>Why AI Agents Get Stuck in Loops (and Quit Too Early)</title><link>https://gabezen.com/blog/llms-control-flow-problem/</link><pubDate>Thu, 18 Dec 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/llms-control-flow-problem/</guid><description>&lt;p&gt;AI agents get stuck in loops, and quit too early, for the same reason: they&amp;rsquo;re bad at deciding how many times to do something. How many files to touch, how long to keep searching, when to stop. A study of transformers learning Collatz arithmetic found the same thing in miniature: 99.7% accuracy, and in over 90% of the failures, the model did the right calculation and miscounted the loop.&lt;/p&gt;&#10;&lt;h2 id="why-do-llms-struggle-with-math"&gt;Why do LLMs struggle with math?&lt;/h2&gt;&#10;&lt;p&gt;There&amp;rsquo;s a common narrative: &amp;ldquo;LLMs are bad at math.&amp;rdquo; People point to failures on arithmetic and conclude that language models lack numerical reasoning.&lt;/p&gt;</description></item><item><title>Is AI Making Developers Worse? Yes, and They're Worth More</title><link>https://gabezen.com/blog/ai-agents-force-multipliers/</link><pubDate>Thu, 18 Sep 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-agents-force-multipliers/</guid><description>&lt;p&gt;&lt;strong&gt;Unpopular opinion: AI coding agents are making developers worse, and I love it.&lt;/strong&gt;&lt;/p&gt;&#10;&lt;p&gt;I can&amp;rsquo;t imagine coding without AI agents anymore. Honestly, I&amp;rsquo;d be bored out of my mind. Yes, I love playing with these tools, and if I lose some low-level skills along the way, so be it. I&amp;rsquo;ve lost plenty of skills before (who remembers Arduino assembly or advanced calculus from university?).&lt;/p&gt;&#10;&lt;p&gt;At the same time, &lt;strong&gt;developers are becoming more valuable.&lt;/strong&gt;&lt;/p&gt;</description></item><item><title>How Law Firms Are Using AI: Notes From a Legal Conference</title><link>https://gabezen.com/blog/ai-adoption-legal-tech/</link><pubDate>Fri, 18 Jul 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-adoption-legal-tech/</guid><description>&lt;p&gt;Law firms are using AI for case research, first drafts of contracts and large document reviews, and a lawyer checks every citation and clause before it reaches a client or a court. That&amp;rsquo;s what I saw at a legal conference, and the surveys mostly back it up.&lt;/p&gt;&#10;&lt;p&gt;I walked in ready for some serious eye-rolls about AI.&lt;/p&gt;&#10;&lt;p&gt;The legal industry has a reputation: conservative, change-averse, &amp;ldquo;we&amp;rsquo;ve always done it this way.&amp;rdquo; I expected skepticism, maybe even hostility toward AI tools.&lt;/p&gt;</description></item><item><title>Panel of Experts Prompting: Templates and What Research Found</title><link>https://gabezen.com/blog/panel-of-experts-prompting/</link><pubDate>Sun, 25 May 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/panel-of-experts-prompting/</guid><description>&lt;p&gt;Most people prompt ChatGPT like it&amp;rsquo;s an intern.&lt;/p&gt;&#10;&lt;p&gt;&amp;ldquo;Write me a marketing email.&amp;rdquo;&#10;&amp;ldquo;Fix this code.&amp;rdquo;&#10;&amp;ldquo;Explain this concept.&amp;rdquo;&lt;/p&gt;&#10;&lt;p&gt;You get much more out of it when you use it like a boardroom of experts.&lt;/p&gt;&#10;&lt;h2 id="the-panel-of-experts-technique"&gt;The panel of experts technique&lt;/h2&gt;&#10;&lt;p&gt;The prompt structure looks like this:&lt;/p&gt;&#10;&lt;blockquote&gt;&#10;&lt;p&gt;&amp;ldquo;I want you to simulate a panel of experts discussing [topic]. The panel includes:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;[Expert 1 with specific credentials]&lt;/li&gt;&#10;&lt;li&gt;[Expert 2 with different perspective]&lt;/li&gt;&#10;&lt;li&gt;[Expert 3 with contrarian view]&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;Each expert should share their perspective on [question], and they should respond to each other&amp;rsquo;s points.&amp;rdquo;&lt;/p&gt;&#10;&lt;/blockquote&gt;</description></item><item><title>Best Backend Framework for Vibe Coding: Why I Pick Django</title><link>https://gabezen.com/blog/django-llm-native-era/</link><pubDate>Sun, 18 May 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/django-llm-native-era/</guid><description>&lt;p&gt;If you&amp;rsquo;re vibe coding a web app, or building one with Claude Code at your side, Django plus HTMX is the stack I&amp;rsquo;d pick. I&amp;rsquo;ve shipped with Flask, Django and FastAPI, and for a small team I still pick Django. I think the reason it works so well with AI comes down to how LLMs learn.&lt;/p&gt;&#10;&lt;h2 id="convention-over-configuration-and-why-ai-loves-it"&gt;Convention Over Configuration (And Why AI Loves It)&lt;/h2&gt;&#10;&lt;p&gt;Django is famously opinionated. Models live in &lt;code&gt;models.py&lt;/code&gt;, views in &lt;code&gt;views.py&lt;/code&gt;, URL patterns in &lt;code&gt;urls.py&lt;/code&gt;, and templates in a &lt;code&gt;templates/&lt;/code&gt; folder next to the app.&lt;/p&gt;</description></item><item><title>Can AI Write Unit Tests? 200 Tests, 15% Caught Real Bugs</title><link>https://gabezen.com/blog/ai-vibe-coding-django/</link><pubDate>Fri, 18 Apr 2025 00:00:00 +0000</pubDate><guid>https://gabezen.com/blog/ai-vibe-coding-django/</guid><description>&lt;p&gt;Yes, AI can write your unit tests, as long as you run them and read them. Over a month Claude wrote about 200 Django tests for me, and roughly 15% failed on the first run because the code was wrong.&lt;/p&gt;&#10;&lt;p&gt;I let AI write both my code and my tests, and the results were better than I expected. Andrej Karpathy gave this style a name in February 2025, when he &lt;a href="https://x.com/karpathy/status/1886192184808149383"&gt;described vibe coding&lt;/a&gt; as the kind where you &amp;ldquo;fully give in to the vibes, embrace exponentials, and forget that the code even exists.&amp;rdquo; There&amp;rsquo;s a common objection to AI-generated code, and it gets louder when the code is vibe coded: &amp;ldquo;How do you know it&amp;rsquo;s correct?&amp;rdquo; The usual answer involves careful code review, which is valid but incomplete.&lt;/p&gt;</description></item></channel></rss>