A Month of Building With AI · a static copy of Robert Boulos's whiteboard, 16 Nov 2026The kit page, the deck prompt and the downloads →
A month of building with AI
Don't trust the agent: gate it, measure it, prove it.
Robert Boulos · 16 Oct – 16 Nov 2026 · for Ray, Daniel and Ivan · every number here was read from the repos, logs and live servers on 16 Nov
The loop behind every project. Each column is one station: what was built, one measured number, and what to steal.
1 · Decide the next move
the next mod
2 · Brief it exactly
one sun in the sky
3 · Build in lanes
Snappy over ACP
4 · Gate the shortcuts
the snappy-rust gate
5 · Prove it live
Feisst + the Builder
6 · Make it a hook
rules that hold
129 of 187
7 parts
77%
308 crates
118 tools
31 hooks
moments he sent a suggested move as is, or edited it
an agent's brief needs, or its launch is refused
of 32 lanes' cost was re-reading context (11 Nov)
866 PRs, 221 merged, in the Rust daemon's first 6 days
on the live Feisst MCP today; 79 on 18 Oct
carry his rules into every Claude Code session
A Claude Code mod that keeps the next message above the prompt: two moves labelled by tactic (Push, Root cause, Close the loop, Unblock, Answer first) and a meter for whether the ask has a check that proves it. A change ships only after a replay of 187 real moments, where Jev scores each card against what he actually typed next.
The top seat, the strongest model, does the thinking; Sonnet agents only execute. A hook refuses a launch whose brief lacks GOAL, FILES, STEPS, PROOF, RETURN, PROGRESS and DO NOT, or that says “investigate” or “figure out”. A stop hook won't let the coordinator end a turn on a failure until it has written the cause, with file and line.
Each lane is a real Claude Code, Codex or Gemini session in its own worktree, hosted over the Agent Client Protocol as a durable per-folder session that streams every tool call, diff and permission ask. 4,839 calls re-read 1,517M tokens, 318k each; polling for gates was 22%. A hook now refuses polls and over-spent resumes.
Lanes write the Rust daemon; a gate written against how agents cut corners judges it. Crates cap at 6,000 lines, modules at 800 (a 400-line file cap was dodged with 77 _part files). Deleting a test fails. The daemon must boot and answer /health in 20 s: PR #307 passed every test, then panicked at start. Mergify merges every green PR.
A Cloudflare Worker that lets Feisst Group staff ask Claude or ChatGPT about simPRO, Jobpac, Navman and BambooHR, scoped to the person, with cards in the chat. Nightly it asks the live server, as every person on the register, what they can see. The Builder turns a staff request into a tool; only a read of the app's own release marks it Live.
A rule in a prompt fades at the next compaction; a hook holds. No builds or disk-wide searches on the MacBook (its load hit 110). One live process per session. Every Stop runs the strictest linters on changed files. Old models can't be used to judge. Judging goes to Jev, a typed model: a yes/no, a choice or a score with a confidence, in 263 ms.
Steal this: most prompt edits are noise. The live prompt's own reruns scored 9–12 of 42; every variant tried scored 8–11. And show a suggester only the human's messages and the agent's last reply, never tool output.
Steal this: an agent sent to “go figure it out” burns hours while you guess. If you can't write its steps, the problem isn't solved yet. This hook refused the research agents for this board too.
Steal this: treat a launch as answered when the agent accepts it, not when its turn ends. A 5-minute block made callers relaunch healthy 28-minute turns as failures.
Steal this: write each gate rule against a shortcut you actually saw, and keep the incident next to it. When it fires, the rule explains itself.
Steal this: done is a fresh read of the real thing after the act, never the agent's report. Switching on the Builder's live probe exposed two bugs in the probe itself the same night.
Steal this: a judge should return a number you can set a threshold on, not prose to parse. Don't ask a second LLM to check the first. Put the date and the incident on every rule.
Where it runs
MacBook · Robert and the top seat. No builds, tests or disk-wide searches here: a hook sends them away.
Mac mini · the workhorse: lanes, and a 60-second sweep that alarms on disk, stuck locks, hung tests and drifted lanes.
Hetzner hz1–hz4 · every PR gate, the Mergify queue (4 drafts checked at once), the Builder's agent.
Cloudflare + Xano · the doors: Feisst's MCP and app, the Builder, Snappy's MCP. Xano holds the data and control plane.
Lessons that became rules
6 Nov · A test passes only when it proves everything that was asked, against the real source. Making a check fail on purpose proves only that it can fail.
Success is the default lie. An empty grep, a missing file or a refused API call reads as zero. Distrust any check that passes on nothing.
26 Oct · Eight lanes wrote TypeScript beside the Rust migration. Now a pre-commit gate refuses any .ts, and each offending branch owed two TS lines deleted per line it added.
11 Nov · The daemon decides, the app draws. Every state, refusal and reason comes from the server; the app formats and lays out. Test: would Xano do this, or the React component?
A value you don't hold is “unavailable” with one sentence saying why. Never blank, never invented, never “sign in again”.
Rows, not code. In the companion app, vibes, video models, provider routes and GPU rentals are database tables: a vibe edit reaches the app in a minute, no deploy.
12 Nov · Read the provider's docs and run its defaults before any paid run. A rented H200 and four wasted jobs taught this one.
When the human names the workflow (“launch subagents”, “/loop”), the workflow is the spec, not the model's idea of what's faster.
The month in numbers
Repo16 Oct – 16 Nov
snappy-runner (TS, retiring)817 PRs merged
snappy-rust (since 10 Nov)221 PRs merged
snappy-swift (the Mac app)120 PRs merged
snappy-kernel (skills)90 PRs merged
feisst (client MCP)799 commits
companion app (Xano)245 commits
Show it live tomorrow
  • the next band above a Claude Code prompt
  • a Feisst question in Claude Desktop, with its card
  • a Builder request moving to Live
  • the Mergify queue taking green PRs on snappy-rust