Clankers

In four words

I asked ChatGPT: "How best to describe me in 4 words". Its reply, word for word:

If I had to compress how you come across into four words:

Curious. Ambitious. Analytical. Restless.

• Curious — you explore everything from AI and software to history, psychology, investing, travel, and even nursery rhymes.
• Ambitious — you’re consistently thinking about the next level: career, wealth, projects, family, lifestyle.
• Analytical — you naturally break things into systems, numbers, trade-offs, and scenarios.
• Restless — probably the most distinctive one. You rarely seem satisfied with “good enough”; there’s usually a “but what if we did this better?” lurking underneath.

If I were allowed a fifth, I'd pick builder.

The workbench

I work in Claude Code. My config is public: claude-config.

One model does the thinking. It decides, designs, reviews and talks to me. Cheaper models do the legwork:

• Scouts (Haiku) look things up, one question each.
• Verifiers (Haiku) run the checks and report back. They never fix anything.
• Builders (Sonnet) write code from a written brief, inside a fixed scope.
• A browser agent (Sonnet) checks pages in my real Chrome.

Builders never commit. I do, after the change has been walked through with me.

Skills I wrote

• own-it: walks me through every change until I could defend it myself.
• orchestrate: the delegation rules above.
• pending-decisions: lays out every decision waiting on me, with enough context to decide on the spot.
• review-mr: reviews a pull request the way its author would: runs the tests, reproduces the steps, and posts nothing until I say go.
• Also: git-commit, write-goal.

Skills by others I use

stop-slop, web-design-guidelines, frontend-design, impeccable.

Hooks

A read cap that keeps the thinking model's context small, session hooks that feed Membral, and a status line.

Membral, the memory

Membral (source) saves my coding sessions from Claude Code, Codex, opencode, Copilot and Antigravity into one local store. It pulls out the decisions, learnings and outcomes and makes them searchable, so a new session starts with what the last ones found. A sweep picks up new sessions every 20 minutes.

The bots

1. DropDaily at 20:30I approve

   Mines my day in Membral for one thing I made or fixed, drafts a post with an image and sends it to me on Telegram. Nothing goes live until I reply "go". Then it commits and pushes to this site. It also sends me a morning brief at 08:30, a taste check on Sundays, and watches its own health twice a day.

2. StarEvery 30 minutesI approve

   Looks for people talking about Membral and drafts replies for me.

3. CoderTwice an hourI approve

   Picks a queued task on Membral and opens a draft pull request. It never touches main. I merge.

4. ReviewerTwice an hourI approve

   Reviews the coder's pull requests against a written gate.

Staytion

Staytion is my paying-guest coliving business in Mumbai. Two clankers keep it running, alongside an Android app for property, tenant, rent and electricity tracking, with WhatsApp built in.

1. OpenClaw agentAround the clockruns on its own

   Runs on a small VPS as an OpenClaw agent on a dedicated WhatsApp number. It sends the reminders, gives us an interface to the Staytion backend, and handles the admin, from agreement signing to rent reminders.

2. SallyEvery 10 minutesruns on its own

   Sally is our sales agent. She does the selling, working to fill empty beds. She can't mark a bed sold; only a separate reconcile step can. Her messages go out only after approval, checked every minute.

Want to compare setups? surajtripathi793@gmail.com