b.akery
Own the outcome
without writing the code.
I don’t write the code; I write the intent and judge the output. You make the rules; b.akery makes your agent remember and enforce them. Corrections stick, so you stop repeating yourself. The more you use it, the better it gets. b.akery isn’t another AI. It’s a better way to work together.
Everything on this page is the method demonstrating itself: the speed, the reasoning, the polish. This page included.
- 01 Commit stamps
- how fast it shipped — timestamped, not claimed
- 02 A shipped storefront
- operate the real content-and-commerce flow
- 03 A call console
- ask for a call from a voice agent I built
- 04 How it works
- four failure modes, four gates — try one live
- 05 Kieran’s Forest
- a game a different AI built under the same rules — play it
Five exhibits today. The room fills up: the argued decisions, the front-office system.
This site, timestamped.
The honest way to prove speed is the record. Cloudflare writes the deploy log, not me, and the public git history has to agree with it. So here is this exact page, from nothing to live.
- 2026-07-15first commitAn empty repository. No design, no copy, no pipeline.
- ~1 dayto live on the apexFirst commit to jamesdnichols.com serving real traffic, 2026-07-16. A Cloudflare deploy stamp, not a claim.
- ~19.1 daysto built out263 commits and counting — the lander, the guides, the calculators, the measurement layer, this room.
Every figure here is read from the git log at build time, so it can only ever be current. git + Cloudflare stamps.
A storefront, rebuilt so you can click it.
Content plus commerce: a park page, a guide, a checkout. Operate the flow — this is the outcome the method ships, not a screenshot of it.
idea → monetized content, ~6 weeks
Astro · Cloudflare Workers + D1 · Shopify
Atlas · national parks
Arches National Park
Red rock, high desert, and the most concentrated set of natural arches on earth. Plan the trails, the light, and the crowds before you go.
The Arches Field Guide
Trail-by-trail, where to park, and the best hour for photographs.
$12
Facsimile. Rebuilt from scratch in Astro for this page — not the live store, no real checkout, and the park and guide are illustrative. The real Atlas sold national-park guides on Shopify.
Ask a question. A system I built calls you back.
Clara answers the phones for home-service contractors. The strongest kind of proof is the one you can talk to — so leave a number and she rings you.
idea → Clara on a live line, ~2 weeks (git: 2026-04-05 → 04-18)
Retell AI · SIP telephony · Cloudflare Workers + D1
Clara answers the phones for home-service contractors — books jobs, quotes windows, never sleeps. Ask her anything an HVAC customer would, and she calls you back.
Runs on TradeFront’s registered calling stack — A2P 10DLC brand and TCPA consent, the part almost every agent demo skips.
Live. Leave a number and Clara dials it once, on a recorded line, under TradeFront's registered A2P 10DLC and TCPA consent trail. Outbound-only and opt-in by design — no open number to abuse, one call per number per day, hard daily cap.
What b.akery can’t fix.
Your agent will still have bad days. Three you’ll recognize:
- Anything with numbers
- Claude calling it a night in the middle of the afternoon? Math that doesn’t add up?b.akery bans invented figures. Numbers and dates come from the record, and anything untraceable doesn’t ship.
- Sycophantic theater
- Suspicious enthusiasm, or mildly infuriating micromanagement?The rules set the tone once: no flattery, no theater. The checks don’t care how confident the answer sounded.
- Sudden lobotomy
- Is your agent gaslighting you, or did it genuinely forget the key details?It doesn’t get to rely on memory. Every session starts by rereading the rules, the corrections, and where the work left off.
We can’t fix the shortcomings of current AI. We can make them rarer, cheaper, and less maddening. The best part: as the models improve, b.akery improves with them, and bigger, longer, more autonomous work becomes a benefit, not a burden.
How b.akery works.
You've used ChatGPT or Claude. It wrote confident garbage, invented a statistic, forgot everything by Tuesday, and said "done" about things that were visibly broken. Same models here — different results. The difference is a system built around four specific failure modes.
1 It forgets everything.
Every session starts by rereading a standing contract: what is true, what is in flight, how to behave. Next to the facts sits a stranger thing — a written list of what the model will plausibly get wrong, with dated corrections it is told not to undo. The amnesia is assumed, then engineered around.
"The launch-kit 470k/400k figures are SUPERSEDED — do not 'correct' surfaces back to them."
A real entry, written to stop a future AI session from "helpfully" reverting a corrected fact. It worked.
2 It invents numbers.
No figure ships from the model's mouth. Every public number traces to an analysis document carrying the query that reproduces it and the base it was measured on — and where the model does arithmetic, one math file holds every formula, derived in comments, with tests that assert hand-computed values only.
This site's flashiest stat was a 19.5x ROAS. Held to the light: it counted only $122,830 of the $414,273 actually spent — a selection artifact. It got demoted to the fine print, and the page now leads with 9.2x on every dollar. The method downgraded its own best number. The demotion is itself on the record: a dated addendum in the analysis doc, with the old pair still readable in the fine print, labeled.
// engine.mjs: "No constant in this file comes from a language model."
3 It says "done" when it isn't.
Rules the model "should follow" are worthless — so here they're checks that fail the build. Typecheck, arithmetic tests, contrast math, and a copy gate that bans the builder's own jargon from ever reaching a buyer. And every checker must first catch fifteen planted errors before it is allowed to certify anything: a gate that can't see known-bad pages doesn't get an opinion.
$ try the gates — the same regexes that gate this page's deploys, imported from the same file
Verdicts print here, exactly as the build prints them.
For the developers — the eaten-space check, as shipped:[A-Za-z0-9—.?!,:;]<(strong|em|b|i|a|span)\b[^>]*>[A-Za-z0-9] · it reads the raw HTML because every other scanner strips tags to spaces first and literally cannot see this bug.
4 A green checkmark isn't the truth.
Automated checks render every changed page at phone and desktop width, light and dark, and assert computed styles. Then the part no tool can do: a commit that touches design is blocked unless a dated "I actually looked" entry ships with it. The hook is honest about its limit — it checks the log moved; the looking is the point.
"The dev server was serving an error page. vision-verify printed 'ok'. Only looking caught it."
Three separate entries where every mechanical gate was green and a human eye still caught the real bug. That log is why the gates are trusted: they're allowed to be wrong, on record.
Not a one-off.
The same discipline runs in five systems, and every rule below was born from a real incident in that repo — one per month, months before this page existed.
- MAR 2026 · margin pipelineA one-character deploy flag wiped production credentials. The rule went into the log the same week, and a machine tripwire has re-checked it on every scheduled sweep since.
- APR 2026 · voice agentA test harness phones the AI receptionist and scores the transcript. The failing report stayed checked into the repo until the fix shipped, next day.
- MAY 2026 · treasuryWhere money is counted, the model never does arithmetic: it writes the queries, the database computes, and ingestion ships with run-twice idempotency tests.
- JUN 2026 · storefrontA filled button rendered green-on-green (invisible) and passed every lint. Verification moved to the computed render — the rule this page’s own gate descends from.
- JUL 2026 · dialectic engineOn its first live run, agents “succeeded” while returning auth errors. Hardened the same afternoon: agent-reported errors hard-fail, never salvaged into output.
The model isn't smarter. Its failure modes are made non-recurring.
The claim ends where the gates end: these systems, run this way, by an owner who doesn't write code. Every escaped bug becomes a permanent gate, the same day. That's the whole method, and it's what you're reading right now.
One pasted prompt. A playable game. Every correction on the record.
A different AI this time: ChatGPT, handed the b.akery starter kit, built this from one sentence typed on a walk to a coffee shop. What it left behind matters more than how fast it went.
first phone-playable demo → shipped: 3 user turns, 2 design revisions
ChatGPT · cloud work session · GitHub
an isometric forest, playable now
Kieran’s Forest
Play Kieran’s Forest ↗opens in a new tab · built phone-first
$ the build, from the record
- the pasteOne install prompt, pasted into a fresh ChatGPT session from my phone. No laptop, no code editor: the whole build ran in the AI’s own cloud workspace. The conversation ran thirty-one messages; I typed five of them.
- two verdictsTwice I played a demo and said fix it first. Each verdict became a dated ledger entry, new standing rules, and a test that runs on every release since.
- the self-auditThen the agent audited its own build: 727 lines, adversarial. It ruled its own earlier claim of a “complete playable demo” overstated. That ruling is why this exhibit’s numbers start at the first demo I could open on my phone.
- one rule went upstreamThe audit came with a scope report; its proposed boundary is now rule fifteen of the standing contract every future build rereads.
- the one manual actThe only thing done by hand in the whole build: creating the GitHub repo that keeps the transcript, the audit, and the harness files. The workspace was disposable. The record is the asset.
- the recordTranscript and audit: private today, published when that repo opens. Until then, this log and the game itself are the public evidence.
Live and public: the link opens the real game, hosted where ChatGPT built it. Built phone-first, so play it on your phone. The walk and the one-line description are my memory; every count in the log is recounted from the saved conversation.
Want this pointed at your business?
Leave an email and I'll send the step-by-step approach I use to find lost margin: the same method as a paid scan, minus the labor. No spam, no hard sell.
Rather build it yourself?
Get the free starter kit at b.akery.org →