Tech Notes to Self

Writer’s Brief (Outline) This section is a handoff for a writer. The Hook I had a 25-page playbook for building authority (InfluenceOS) and months of call transcripts where real people reacted to my ideas. I asked Claude to turn those two things into a single system that tracks what’s actually working. Main Points to Cover The core technique: instead of guessing what your “narrative spine” should be, analyze actual conversations to find what people already repeat back to you The surprising part: the phrases I thought were strongest weren’t the ones that got echoed. “Just ask Claude” beat everything. “Ego death” emerged independently in multiple sessions without me pushing it. The practical result: a slash command that reads a living data file, knows who you’ve talked to, what’s resonating, and what assets are missing. Not a content calendar. An authority tracker. The Angle Personal discovery. I had all this raw material scattered across transcripts and docs and a PDF playbook I liked. The interesting part was what happened when I combined them: the evidence-based approach revealed gaps I was blind to. ...

Writer’s Brief (Outline) This section is a handoff for a writer - what the post is about and what to cover. The Hook What if your AI assistant kept working after you closed the laptop? Main Points to Cover The pattern: launchd (macOS scheduler) + claude -p (headless Claude) = agents that run on autopilot The first implementation: Air Traffic Control - an inbox router that sorts your brain dumps into the right files, twice a day, without you touching anything The bigger idea: “capture is cheap, sorting is free” - you just throw stuff into a pile and the system handles the rest The Angle Personal discovery. Built a bunch of Claude Code skills that require manual invocation, then realized the boring ones could just… run themselves. The moment it clicked was seeing the Chief of Staff inbox get a routing summary from an agent that ran while I was making coffee. ...

Step-by-Step Tutorial What This Does Turns a basic content dashboard into a full lifecycle tracker. Posts/comments/DMs flow through status stages (draft → approved → posted), get grouped by week, and show up in filterable pills. Plus you get a “this week” stats row showing what actually shipped. The Prompt Add status flow tracking to the Britt HQ dashboard: - Posts: draft → approved → posted - Engagement: pending → approved → commented - Outreach: to_send → approved → sent → connected Add filter pills for each status, weekly tracking for all three columns, and reorganize the layout so items needing review are in row 1 and "this week" stats are in row 2. What Claude Does Defines status flows for each content type Writes ISO week calculation function to group items by week Adds filter pill UI with click handlers Creates status badge components with color coding Builds weekly stats cards (counts items that reached final status this week) Reorganizes dashboard: actionable items top row, progress metrics bottom row Tests status transitions by changing markdown files The tricky part: engagement uses checkbox syntax (- [x]), outreach uses section headers (## Sent), but posts use inline tags ([draft]). Had to normalize all three into a unified status model.

Step-by-Step Tutorial What This Does You type /chief-of-staff and Claude reads 15+ context files (your inbox, recent meetings, content drafts, social queue, experiments), synthesizes everything into 10 prioritized questions tagged to your goals, and generates an HTML page that walks you through them one at a time with quick-action buttons and a progress bar. You respond in the browser, copy everything back, and Claude processes your answers. The Prompt /chief-of-staff [Claude generates a wall-of-text check-in with 10 items] "Can you build me an HTML page that walks through these one at a time? I want to respond to each item in the browser and then paste everything back to you." What Claude Does Creates an HTML file with dark theme and monospace font Shows one card at a time (no scrolling through everything) Each card has: priority fires, title, context paragraph, goal tag, question, quick-action chip buttons, textarea, skip button Progress bar tracks where you are (3/10, 7/10) At the end: compiles all responses into a markdown block with “Copy to Clipboard” button You paste back, Claude updates all your state files

Step-by-Step Tutorial What This Does Reduces your CLAUDE.md configuration file from 800+ lines to ~100 lines by extracting reference material to external files. Result: 85% token reduction, faster responses, and the weird satisfaction of watching a config file go on a diet. The Prompt can you analyze my claude.md file based on these ideas? and suggest edits? [link to Medium article: "Your CLAUDE.md Is Probably Wrong: 7 Mistakes Boris Cherny Never Makes"] What Claude Does Finds TWO identical 789-line CLAUDE.md files (one global, one project-level) both loading in every conversation Calculates you’re burning ~17,000-20,000 tokens before any actual prompt Backs up both files with dated suffixes Creates ~/.claude/references/ directory Extracts Netlify docs, API configs, directory maps, and client patterns to 6 separate reference files Rewrites main CLAUDE.md to 107 lines (keeps behavioral rules, removes encyclopedic content) Deletes the duplicate file Adds missing Workflow and Verification sections from the article’s framework Removes hardcoded API keys for security Prunes stale content (task items from months ago, outdated goals) Final score: From 1,578 total lines to 107 lines. ~1,200 tokens vs ~20,000 tokens.

Writer’s Brief The Hook I discovered that 15 lines of my Claude Code memory file were being silently dropped every single session. No error, no warning. Just gone. Main Points to Cover How Claude Code’s memory system actually works under the hood (it’s not a system prompt, it’s injected as a “system-reminder” in user messages with a disclaimer that says your instructions “may or may not be relevant”) The 200-line hard cap on MEMORY.md that nobody talks about, and the compliance data showing monolith files at 92% vs modular rules at 96% The refactor: splitting a 197-line monolith into 7 focused rule files, compressing 215 lines of memory to 78, and adding a hook that force-loads learnings every session The Angle Personal discovery / problem solved. “I built this elaborate memory system for Claude and it was working… mostly. Then I looked under the hood.” ...

Writer’s Brief (Outline) This section is a handoff for a writer — what the post is about and what to cover. The Hook I have a file that tells Claude what day it is, and Claude still gets it wrong. Main Points to Cover The date is RIGHT THERE in the system context — Claude just doesn’t look at it I already had instructions telling Claude to check the date — they weren’t working The “fix” is making the instructions louder and more aggressive (uncertain if it’ll work) The Angle The absurdity of having correct instructions that get ignored. The meta-humor of yelling at a text file hoping the AI will listen harder. ...

Step-by-Step Tutorial What This Does Removes sensitive files from a public GitHub repo AND scrubs them from the entire git history so they can’t be recovered from old commits. The Prompt I need to clean up the public GitHub repo - found some private data that shouldn't be there. Remove it and make sure it's gone from the git history too. What Claude Does Creates an allowlist-style .gitignore (ignore everything except deploy files) Removes files from current git tracking Installs BFG Repo-Cleaner Clones a mirror of the repo Runs BFG to scrub files from entire history Force pushes the cleaned history to GitHub Lists all other public repos and makes them private (because why were 28 repos public anyway?)