<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>context-engineering on S Anand</title>
    <link>https://www.s-anand.net/blog/tag/context-engineering/</link>
    <description>Recent content in context-engineering on S Anand</description>
    <generator>Hugo -- 0.165.0</generator>
    <language>en-us</language>
    <lastBuildDate>Sat, 16 May 2026 22:24:32 +0800</lastBuildDate>
    <atom:link href="https://www.s-anand.net/blog/tag/context-engineering/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>How I use Local MCP</title>
      <link>https://www.s-anand.net/blog/how-i-use-local-mcp/</link>
      <pubDate>Sat, 16 May 2026 22:24:32 +0800</pubDate>
      <guid>https://www.s-anand.net/blog/how-i-use-local-mcp/</guid>
      <description>&lt;p&gt;&lt;img loading=&#34;lazy&#34; src=&#34;https://files.s-anand.net/images/2026-05-16-how-i-use-local-mcp.avif&#34;&gt;&lt;/p&gt;
&lt;p&gt;I&amp;rsquo;d love for Claude or ChatGPT to answer questions like:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;What meetings am I not setting up that I really should be?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;or:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Based on my activities since 9 May 2026, what should I blog about?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;or:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Who in my professional life most deserves an unreasonable gesture?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;From data. My files, emails, calendar, contacts, transcripts, blogs, notes, code, browsing history, logs, random Markdown files I forgot I wrote.&lt;/p&gt;
&lt;p&gt;Hence, a &lt;a href=&#34;https://modelcontextprotocol.io/&#34;&gt;Local MCP&lt;/a&gt;.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;My Local MCP server exposes one tool: &lt;code&gt;bash&lt;/code&gt;.&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-python&#34; data-lang=&#34;python&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;nd&#34;&gt;@mcp.tool&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;()&lt;/span&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;async&lt;/span&gt; &lt;span class=&#34;k&#34;&gt;def&lt;/span&gt; &lt;span class=&#34;nf&#34;&gt;bash&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;(&lt;/span&gt;&lt;span class=&#34;n&#34;&gt;commands&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;:&lt;/span&gt; &lt;span class=&#34;nb&#34;&gt;str&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;,&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;ctx&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;:&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;Context&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;)&lt;/span&gt; &lt;span class=&#34;o&#34;&gt;-&amp;gt;&lt;/span&gt; &lt;span class=&#34;nb&#34;&gt;str&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;:&lt;/span&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;    &lt;span class=&#34;s2&#34;&gt;&amp;#34;&amp;#34;&amp;#34;Runs multiline bash script.&amp;#34;&amp;#34;&amp;#34;&lt;/span&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;    &lt;span class=&#34;n&#34;&gt;result&lt;/span&gt; &lt;span class=&#34;o&#34;&gt;=&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;subprocess&lt;/span&gt;&lt;span class=&#34;o&#34;&gt;.&lt;/span&gt;&lt;span class=&#34;n&#34;&gt;run&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;(&lt;/span&gt;&lt;span class=&#34;n&#34;&gt;commands&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;,&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;shell&lt;/span&gt;&lt;span class=&#34;o&#34;&gt;=&lt;/span&gt;&lt;span class=&#34;kc&#34;&gt;True&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;,&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;executable&lt;/span&gt;&lt;span class=&#34;o&#34;&gt;=&lt;/span&gt;&lt;span class=&#34;s2&#34;&gt;&amp;#34;/bin/bash&amp;#34;&lt;/span&gt;&lt;span class=&#34;p&#34;&gt;)&lt;/span&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;    &lt;span class=&#34;k&#34;&gt;return&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;result&lt;/span&gt;&lt;span class=&#34;o&#34;&gt;.&lt;/span&gt;&lt;span class=&#34;n&#34;&gt;stdout&lt;/span&gt; &lt;span class=&#34;o&#34;&gt;+&lt;/span&gt; &lt;span class=&#34;n&#34;&gt;result&lt;/span&gt;&lt;span class=&#34;o&#34;&gt;.&lt;/span&gt;&lt;span class=&#34;n&#34;&gt;stderr&lt;/span&gt;
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;That&amp;rsquo;s it. No vector database. No UI. No custom connectors. No &amp;ldquo;AI knowledge platform.&amp;rdquo;&lt;/p&gt;
&lt;p&gt;Just: &lt;strong&gt;run shell commands on my machine&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;I run this locally, expose it online (which is slightly scary), and give Claude and ChatGPT this &lt;a href=&#34;https://www.s-anand.net/blog/prompts/fragments/#local-mcp&#34;&gt;prompt fragment&lt;/a&gt;:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-markdown&#34; data-lang=&#34;markdown&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Local MCP runs bash and exposes:
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;-&lt;/span&gt; ~/code/talks/README.md - talk transcripts, slides
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;-&lt;/span&gt; ~/code/blog/description.md - 20K files, 5K posts. Search for &amp;#34;- llm&amp;#34; for AI-related posts.
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;... (etc.)
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;sb&#34;&gt;`gws`&lt;/span&gt; can access email, calendar, etc.
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;In one shot, this gives EVERYTHING I have to the agents.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;A common use is meeting prep.&lt;/strong&gt;&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;You are a brilliant, brutally honest Chief of Staff. You have full access via Local MCP bash tool to calendar, emails, and past transcripts. Produce a briefing card for each substantive external meeting today.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;It checks the calendar via &lt;code&gt;gws&lt;/code&gt;. It searches my transcripts. My notes. My &lt;a href=&#34;https://www.s-anand.net/blog/ai-advice/&#34;&gt;AI advice&lt;/a&gt;. Then gives me a briefing card with everything I need.&lt;/p&gt;
&lt;p&gt;I can&amp;rsquo;t do this by uploading files manually. The context is not one file: it&amp;rsquo;s scattered all over.&lt;/p&gt;
&lt;p&gt;A human assistant could do this. But agents are faster, cheaper, and I trust them more.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;Another common use is relationship intelligence.&lt;/strong&gt;&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;What meetings am I not setting up that I really should be?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Claude scans transcripts, contacts, emails, and recent activity to find people I should speak.&lt;/p&gt;
&lt;p&gt;This is where Local MCP is different from a file upload.&lt;/p&gt;
&lt;p&gt;In a file upload, I can ask &amp;ldquo;Where is X?&amp;rdquo;.&lt;/p&gt;
&lt;p&gt;Here, I&amp;rsquo;m asking &amp;ldquo;What am I missing?&amp;rdquo; and the answer depends on recency, relationship history, frequency, how conversations felt, unresolved actions, and so much more.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;A third use is mining my own work.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;I used Local MCP to ask what I should blog about. It scanned all my content and found themes I haven&amp;rsquo;t really thought about, like:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;a href=&#34;https://www.s-anand.net/blog/google-meet-captions-local-transcript-recorder/&#34;&gt;Google Meet captions&lt;/a&gt; - a code commit I recently made. I wrote about it.&lt;/li&gt;
&lt;li&gt;&lt;a href=&#34;https://talks.s-anand.net/2026-05-15-gramener-all-hands/&#34;&gt;Agents are the new software&lt;/a&gt; - a theme I&amp;rsquo;ve been talking a lot about. I wrote about it.&lt;/li&gt;
&lt;li&gt;Local MCPs - that&amp;rsquo;s this post&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&amp;hellip; and half a dozen topics I should be writing about soon.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;A fourth use is business research.&lt;/strong&gt;&lt;/p&gt;
&lt;p&gt;I have transcripts from sales calls and client conversations. I don&amp;rsquo;t attend all of them. But Local MCP can.&lt;/p&gt;
&lt;p&gt;I can ask:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Which client needs have we heard repeatedly but not converted into demos?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;or:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Which solutions have we pitched to one client that another client has explicitly asked for?&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;This is beyond a CRM search. A HubSpot search finds what people typed in. This finds what people actually said.&lt;/p&gt;
&lt;p&gt;Then an email search finds if they acted on it. Calendar search finds what we spent time on.&lt;/p&gt;
&lt;p&gt;Across these, I find opportunities that no single system has.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;BUT:&lt;/strong&gt; this is not safe by default. A bash MCP server can delete my files, run commands, read my browser sessions, send emails &lt;code&gt;gws&lt;/code&gt;, and all sorts of risky things.&lt;/p&gt;
&lt;p&gt;So I monitor the commands like a hawk, and give it fairly controlled access, and only when I&amp;rsquo;m actually running one of these use-cases.&lt;/p&gt;
&lt;p&gt;I tried OAuth but setting up Auth0, dynamic client registration, callback URLs, scopes, ChatGPT connector errors, &amp;hellip; I gave up.&lt;/p&gt;
&lt;p&gt;For now, supervised local usage gives me most of the value.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;BUT #2:&lt;/strong&gt; Claude and ChatGPT use Local MCP differently.&lt;/p&gt;
&lt;p&gt;Claude uses it beautifully. Smooth. No mistakes. References memory.&lt;/p&gt;
&lt;p&gt;ChatGPT is more restrictive. No chat memory accessed, nor saved. Keeps asking for permissions.&lt;/p&gt;
&lt;p&gt;So I use ChatGPT less for Local MCP-heavy tasks. But ChatGPT is rigorous. When I want structured analysis, exhaustive lists, or better verification discipline, it is useful.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;Local MCP is powerful because it lets AI &lt;em&gt;use all systems I have access to&lt;/em&gt;:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;local files - across Dropbox, Google Drive, my notes, blog posts, transcripts, slides, &amp;hellip;&lt;/li&gt;
&lt;li&gt;code - not just reading, but running, rewriting, and generatig&lt;/li&gt;
&lt;li&gt;email, calendar, contacts&lt;/li&gt;
&lt;li&gt;browser history&lt;/li&gt;
&lt;li&gt;shell tools - which can be used to access even more system&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Local MCP invites Claude / ChatGPT as a real assistant into my laptop.&lt;/p&gt;
&lt;p&gt;And into my 2,700-line TODO archive.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;You probably shouldn&amp;rsquo;t expose a bash tool to an AI. But note the direction I&amp;rsquo;m going with this:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;If your work and transactions are agent-readable, your past work compounds.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;If they are trapped in apps, screenshots, and memory, your AI has amnesia.&lt;/li&gt;
&lt;/ul&gt;
</description>
    </item>
    <item>
      <title>Google Meet captions as a local transcript recorder</title>
      <link>https://www.s-anand.net/blog/google-meet-captions-local-transcript-recorder/</link>
      <pubDate>Sat, 16 May 2026 13:32:06 +0800</pubDate>
      <guid>https://www.s-anand.net/blog/google-meet-captions-local-transcript-recorder/</guid>
      <description>&lt;p&gt;&lt;img loading=&#34;lazy&#34; src=&#34;https://files.s-anand.net/images/2026-05-15-google-meet-captions-tool.avif&#34;&gt;&lt;/p&gt;
&lt;p&gt;I&amp;rsquo;m a man of simple needs. All I want is: when I&amp;rsquo;m on &lt;a href=&#34;https://meet.google.com/&#34;&gt;Google Meet&lt;/a&gt;, I turn on captions. I wanted to click a bookmarklet and save those captions into a local Markdown file. (So that an AI agent can guide me from it.)&lt;/p&gt;
&lt;p&gt;Hence, &lt;a href=&#34;https://tools.s-anand.net/gmeetcaptions/&#34;&gt;Google Meet Captions&lt;/a&gt;. The code is in &lt;a href=&#34;https://github.com/sanand0/tools/tree/main/gmeetcaptions&#34;&gt;&lt;code&gt;gmeetcaptions/&lt;/code&gt;&lt;/a&gt;. Drag the button to your bookmarks bar. Join a Meet. Turn on captions. Click it.&lt;/p&gt;
&lt;p&gt;You get a tiny panel with two buttons: &lt;strong&gt;Copy&lt;/strong&gt; and &lt;strong&gt;Start Recording&lt;/strong&gt;.&lt;/p&gt;
&lt;p&gt;&lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/gmeetcaptions.js&#34;&gt;The bookmarklet&lt;/a&gt; writes this kind of Markdown:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-markdown&#34; data-lang=&#34;markdown&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;gh&#34;&gt;# Meeting title
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;-&lt;/span&gt; **Meeting code**: abc-defg-hij
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;-&lt;/span&gt; **Started**: 5/15/2026, 8:00:00 AM
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;k&#34;&gt;-&lt;/span&gt; **Participants**: Alice, Bob, Carol
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;---
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;gu&#34;&gt;## Alice [0:12]
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Good morning everyone.
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;&lt;span class=&#34;gu&#34;&gt;## Bob [0:18]
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;
&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;Let&amp;#39;s get started with the agenda.
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;That&amp;rsquo;s it. No server. No extension. No login. No API. Just a &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/index.html&#34;&gt;bookmarklet page&lt;/a&gt;, a &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/script.js&#34;&gt;script&lt;/a&gt;, and local browser APIs.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;BUT&lt;/strong&gt;: Google Meet captions are live and unstable.&lt;/p&gt;
&lt;p&gt;A sentence may appear as:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-markdown&#34; data-lang=&#34;markdown&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;mic, so,
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Then a second later become:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-markdown&#34; data-lang=&#34;markdown&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;mic, So that&amp;#39;s a new person. Okay.
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Then become:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-markdown&#34; data-lang=&#34;markdown&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;mic, So that&amp;#39;s a new person. Okay. Hey. oh, but,
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;If I simply append every change, the transcript becomes garbage. So the bookmarklet keeps updating the active speaker turn until it becomes stable. The implementation uses a &lt;a href=&#34;https://developer.mozilla.org/en-US/docs/Web/API/MutationObserver&#34;&gt;&lt;code&gt;MutationObserver&lt;/code&gt;&lt;/a&gt; plus a one-second polling fallback. After four unchanged polls, it treats the turn as final.&lt;/p&gt;
&lt;p&gt;The tests are in &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/gmeetcaptions.test.js&#34;&gt;&lt;code&gt;gmeetcaptions.test.js&lt;/code&gt;&lt;/a&gt;, using an anonymized fixture at &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/__fixtures__/captions-anonymized.html&#34;&gt;&lt;code&gt;__fixtures__/captions-anonymized.html&lt;/code&gt;&lt;/a&gt;.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;strong&gt;BUT #2&lt;/strong&gt;: Google Meet&amp;rsquo;s DOM is not a public API. Class names like &lt;code&gt;.nMcdL&lt;/code&gt;, &lt;code&gt;.NWpY1d&lt;/code&gt;, and &lt;code&gt;.ygicle&lt;/code&gt; can vanish overnight.&lt;/p&gt;
&lt;p&gt;So the scraper first tries semantic and structural selectors:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;code&gt;[role=&amp;quot;region&amp;quot;][aria-label=&amp;quot;Captions&amp;quot;]&lt;/code&gt; for the captions region&lt;/li&gt;
&lt;li&gt;&lt;code&gt;img[data-iml]&lt;/code&gt; and &lt;code&gt;googleusercontent.com&lt;/code&gt; avatars to identify caption items&lt;/li&gt;
&lt;li&gt;the first &lt;code&gt;&amp;lt;span&amp;gt;&lt;/code&gt; as the speaker&lt;/li&gt;
&lt;li&gt;the last non-image &lt;code&gt;&amp;lt;div&amp;gt;&lt;/code&gt; as the caption text&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Only then does it fall back to obfuscated class names. That selector strategy is documented in the &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/README.md&#34;&gt;&lt;code&gt;README&lt;/code&gt;&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Boring, but also the difference between &amp;ldquo;worked once&amp;rdquo; and &amp;ldquo;might work tomorrow.&amp;rdquo;&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;The weirdest was Chrome writing to a &lt;code&gt;.md.crswap&lt;/code&gt; file while recording. The file appears unfinished until I click &lt;strong&gt;Stop Recording&lt;/strong&gt;. Then Chrome finalizes it.&lt;/p&gt;
&lt;p&gt;This is good, actually. It means the browser is safely streaming to a local file via the &lt;a href=&#34;https://developer.mozilla.org/en-US/docs/Web/API/File_System_API&#34;&gt;File System Access API&lt;/a&gt;. But it also means: &lt;strong&gt;stop the recorder before trusting the file&lt;/strong&gt;!&lt;/p&gt;
&lt;p&gt;I captured these bugs and prompts in &lt;a href=&#34;https://github.com/sanand0/tools/blob/main/gmeetcaptions/prompts.md&#34;&gt;&lt;code&gt;prompts.md&lt;/code&gt;&lt;/a&gt;, because future-me will forget. Future-agent, too.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;Why bother? Because transcripts are not the output. They are raw material.&lt;/p&gt;
&lt;p&gt;Once a meeting is Markdown, I can ask agents to extract decisions, questions, follow-ups, contradictions, reusable prompts, and blog ideas. I can diff it. Search it. Commit it. Feed it to another workflow.&lt;/p&gt;
&lt;p&gt;Meetings now become the &amp;ldquo;context&amp;rdquo; in context engineering!&lt;/p&gt;
</description>
    </item>
    <item>
      <title>Local context repositories for AI</title>
      <link>https://www.s-anand.net/blog/local-context-repositories-for-ai/</link>
      <pubDate>Fri, 20 Mar 2026 07:12:47 +0530</pubDate>
      <guid>https://www.s-anand.net/blog/local-context-repositories-for-ai/</guid>
      <description>&lt;p&gt;When people ask me for connections, I share my &lt;a href=&#34;https://www.linkedin.com/mypreferences/d/download-my-data&#34;&gt;LinkedIn data&lt;/a&gt; and ask them to pick.&lt;/p&gt;
&lt;p&gt;This week, three people asked for AI ideas. I shared my local content with AI coding agents and asked them to pick.&lt;/p&gt;
&lt;p&gt;&lt;img loading=&#34;lazy&#34; src=&#34;https://files.s-anand.net/images/2026-03-20-local-context-repositories-for-ai.avif&#34;&gt; &lt;!-- https://gemini.google.com/u/2/app/b90eaeb839f3e493 --&gt;&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;STEP 1: Give access to content&lt;/strong&gt;. I use a &lt;a href=&#34;https://github.com/sanand0/scripts/blob/7e1dc00d7e1fa36a9949a1e061c1b529928cc175/dev.dockerfile&#34;&gt;Dockerfile&lt;/a&gt; and &lt;a href=&#34;https://github.com/sanand0/scripts/blob/7e1dc00d7e1fa36a9949a1e061c1b529928cc175/dev.sh&#34;&gt;script&lt;/a&gt; to isolate coding agents. To give access, I run:&lt;/p&gt;
&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; class=&#34;chroma&#34;&gt;&lt;code class=&#34;language-bash&#34; data-lang=&#34;bash&#34;&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;dev.sh -v /home/sanand/code/blog/:/home/sanand/code/blog/:ro &lt;span class=&#34;se&#34;&gt;\
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;       -v /home/sanand/code/til:/home/sanand/code/til:ro &lt;span class=&#34;se&#34;&gt;\
&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&#34;line&#34;&gt;&lt;span class=&#34;cl&#34;&gt;       -v /home/sanand/Dropbox/notes/transcripts:/home/sanand/Dropbox/notes/transcripts:ro
&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;This gives read-only access to my blog, things I learned, transcripts, and I can add more. (My transcripts are private, the rest are public.)&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;STEP 2: Ask agents to scan content&lt;/strong&gt;. For &lt;a href=&#34;https://github.com/sanand0/talks/blob/c45b0e272708f82be50a1c48f10791bd0fee8e46/2026-03-18-iitm-office-of-institutional-advancement/prompts.md&#34;&gt;example&lt;/a&gt;, I ask it to read:&lt;/p&gt;
&lt;blockquote&gt;
&lt;ul&gt;
&lt;li&gt;Required blog posts related to LLMs &lt;code&gt;/home/sanand/code/blog/&lt;/code&gt; (especially with the category &lt;code&gt;llms&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;Other relevant transcripts &lt;code&gt;/home/sanand/Dropbox/notes/transcripts&lt;/code&gt; (especially extracted AI advice at &lt;code&gt;/home/sanand/Dropbox/notes/transcripts/extracts/ai/&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;Things I learnt at &lt;code&gt;/home/sanand/code/til/&lt;/code&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;
&lt;p&gt;This makes it explicitly aware of the content and can use it to answer questions.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;STEP 3: Help it do better&lt;/strong&gt;. I often add &amp;ldquo;Use sub-agents as required&amp;rdquo;, which reduces the context and lets them run more in parallel. I also point them to &lt;a href=&#34;https://github.com/sanand0/talks/blob/c45b0e272708f82be50a1c48f10791bd0fee8e46/2026-03-18-iitm-academic-council/ideas-post-mortem.md&#34;&gt;post-mortems&lt;/a&gt; for tips on scanning content effectively.&lt;/p&gt;
&lt;p&gt;&lt;strong&gt;STEP 4: Output as JSON&lt;/strong&gt;. JSON lets me write programs to convert to multiple other formats (e.g. HTML, markdown). I specify the fields I want, how I want them filled, and leave the rest to the agent. &lt;a href=&#34;https://github.com/sanand0/talks/blob/c45b0e272708f82be50a1c48f10791bd0fee8e46/2026-03-18-iitm-academic-council/ideas/ideas.json&#34;&gt;Sample output&lt;/a&gt;.&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;This is not a new technique. It&amp;rsquo;s just context engineering, roughly like:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Connecting ChatGPT/Claude/Gemini/ to Dropbox/Google Drive/&amp;hellip; and asking it to read the content.&lt;/li&gt;
&lt;li&gt;Enabling web search and asking them to search online.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;But I can do this (kind of) safely on my local content and I can also teach it how to scan the content - which is a useful learning.&lt;/p&gt;
&lt;p&gt;Next steps:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;Add README.md to each directory on how to scan the content effectively.&lt;/li&gt;
&lt;li&gt;Think about what content repositories I should add&lt;/li&gt;
&lt;li&gt;Explore &lt;em&gt;combining&lt;/em&gt; content repositories cleverly (e.g. &amp;ldquo;Read my blog and apply lessons to my code.&amp;rdquo;)&lt;/li&gt;
&lt;/ol&gt;
</description>
    </item>
  </channel>
</rss>
