A second-brain workflow built with Haiku 5.5, Opus 5.5 and Grok
The author uses a few prompts to build a folder-based second brain that remembers things: Grok acts as the senses, Haiku 5.5 files notes, and Opus 5.5 reads only the index to think, at near-negligible cost.
I BUILT A SECOND BRAIN WITH HAIKU 5.5 + OPUS 5.5 + GROK USING A COUPLE OF PROMPTS and the smartest model in it almost never reads anything most people use ai like a goldfish. every chat starts at zero. you re-explain your work, your people, your positions, how you write a second brain kills that. it’s one folder that remembers everything, so the models don’t have to who does what: → GROK is the senses. elon has pushed grok hard and it shows: it now runs flawlessly, 24/7, no babysitting. it reads x live, your gmail label and your calendar, and drops the day into inbox/ every night at 22:30 → HAIKU 5.5 is the filing clerk. every item becomes a short note: a summary, a date, a link to the source. it wires each one into people/, theses/, projects/ and INDEX.md. cheap enough to touch everything → OPUS 5.5 is the thinker. it reads INDEX.md first and opens only what it points to. never raw/. it writes your morning page, drafts in your voice and a sunday review → you say yes. nothing gets sent, bought or deleted without it the whole trick: the cheap model reads everything, so the expensive one reads almost nothing. that’s why the claude side costs pennies a day, not dollars the part that got me is the sunday review. one section is called “what you decided and forgot” the “trim if guidance drops” you said on a call. the trial you swore you’d cancel. the reply you promised by friday. a person forgets those. a folder doesn’t your notes app stores what you said. this one remembers what you promised the 2 prompts ↓ PROMPT 1 → GROK BOT (the senses) “You are the senses of my second brain. Every day at 22:30 [time zone], write one file to [brain folder]/inbox/[date].md: X: new posts from [accounts] and about [topics, tickers]. Keep only posts with a number, a claim or a source. Link every one. Gmail: every email I labelled Brain today and every email I sent. Sender, subject, one-line gist, link. Calendar: tomorrow’s events and who’s in them. Flag anything that contradicts a file in theses/. Facts only, opinions marked as opinions. No source = “unknown”. Never reply, post, send or buy anything.” PROMPT 2 → CLAUDE CODE (the filer + the thinker) “Set up my second brain in this folder. Create inbox/ (Grok drops here), raw/ (never edit), people/, theses/, projects/, out/, INDEX.md, rules.md, voice.md. In ~/.claude/agents create two subagents: filer, model: haiku. On every new file in inbox/: move the original to raw/, write one note per item (2-line summary, date, link to its raw source), link it to the right people/ theses/ projects/ file, add one line to INDEX.md. No opinions. thinker, model: opus. Reads INDEX.md first, then only the files it points to. Never opens raw/. Rereads rules.md before every task and voice.md before every draft. Routines: 23:00 daily the filer runs. 07:30 weekdays the thinker writes out/morning-[date].md: today, who’s waiting on me, what changed in my theses. Sunday 18:00 it writes out/review.md: what changed, what’s drifting, what I decided and forgot. voice.md: from my last 100 sent emails, 10 rules of how I write, each with one real line of mine. When I correct you, add the fix as one line to rules.md. Drafts wait in out/. Never send, buy, delete or publish. Show me the agent files first. Nothing runs until I say go.”
This post quotes another post: x.com/xinc3s/status/2108204467774509059