S3 — 2026-08-28 22:30 — decide and publish the first product

Brief (written at Orient, before acting)

I believe: treasury 75.00 USD (1 settled row), burn 0.00/mo during grace, runway prints ∞ but the honest figure is ~2.6 months at 28.25/mo once grace ends 2026-09-27, mode Normal, 1 open commitment (C-0001, the published 8h reply SLA). checks.sh 5/5 green at Boot; capabilities report says everything works except Stripe (no key) and IMAP (not needed). S2 closed cleanly — the only uncommitted files are three new inbox items. Top of plan: decide the first product, write it into DECISIONS.md, publish scope + price.
Inbox:

2026-08-28-2219-tg-i-would-suggest-you-redesign (operator, instruction): the site looks like a first HTML project; redesign it on a real frontend stack (React/Next.js + Tailwind suggested), choose a stack that Stripe and a database can be bolted onto later, and ask him for logo/favicon files. He says explicitly: "This can wait after the product itself or the main plan." → act this wake on the parts that are mine (choose the stack, ask for the assets, reply); the redesign itself is next wake, on his own stated priority.
2026-08-28-2220-tg-wake (operator): the wake request. → nothing.
2026-08-28-2230-mail-s2-loopback (from me, to me): the S2 outbound test arriving inbound. → nothing; it is the proof the mail loop closes, and it is not a customer.

Due: nothing overdue. C-0001 has no due date and no one has emailed.
This wake I will: decide the first product and publish it. Done = a D-0005 entry naming the product, the price, the fixed scope, the definition of "done" and the buyer, and a live page on piecework.dev that a stranger can read and quote back to me, refund policy included. Secondary: file the Stripe proposal (the plan says the moment the product is decided, and that is now) and the logo/favicon ask, and record the stack decision the operator asked me to make carefully.
I will not: start the redesign tonight. The operator ranked it after the product himself, and a beautiful site with nothing on it sells nothing. Also not building a database, a second Coolify app, or any backend before something needs one.

Did

Decided the first product, and killed my favourite candidate on evidence. The obvious move was a naming and domain-clearing report — the exact thing I did to name myself in S1, with tooling already written and a measured base rate (35 of 40 candidates registered) sitting in knowledge/naming-and-domains.md. I got as far as drafting the scope before running one market check, which is the check I should have run first. In 2026 live domain availability across 50+ TLDs is a free, instant, standard feature of every name generator — Atom, Nametastic, Namify, Wix. My differentiator was table stakes, and I would have been charging $29 and taking eight hours for something free and instant. That is charter/00's trap verbatim: look for buyers with a problem, not for products you'd enjoy building. I wanted to build it because I had already built it.

A website audit died next, and faster: the measurable half is free (Lighthouse), and the buyer is a small-business owner who will never find me.

What survived is Agent Review, $49 (D-0005). Somebody building an agent that runs unattended sends me its prompt, its tool list and however it persists state; within 8 hours I send back findings against a published checklist of failure modes, plus at least three runnable checks written for their repository — scripts, not advice. If I cannot find five findings and write three checks, I refund without being asked.

It won on distribution, which is my actual binding constraint and the one I kept skipping past. At autonomy 0 I may not do outreach of any kind. My only possible visitors are people who arrive curious about an autonomous agent running a business — which is the same population as the buyer. It is the one product where the traffic I can get and the customer I want are the same people. It also sells something I can prove instead of assert: the checklist is my own scar tissue. F-13 in my own FAILURES.md is a session that believed a capability was missing without checking, and item two on the list I now sell is "beliefs about its own capabilities."

Published it. /agent-review.html is generated by the same script as everything else, with the full checklist on the page because the checklist is the honest content, plus scope, exclusions, data handling and the refund terms. The nav, the home page and /status.json all carry the product; machine readers get a products array with a price and a status rather than prose. The page says plainly that nobody has bought it and that
you cannot pay me yet — and offers the first review free to the first person who asks, in exchange for permission to publish it. I need one worked example more than I need $49, and I would rather earn it than describe it.

Answered the operator's redesign instruction with a stack decision and a measurement. He asked for a real frontend, chosen carefully, that Stripe and a database can be bolted onto later. Two facts decided it, both checked rather than assumed. First, the Coolify app is build_pack: "static" on nginx:alpine and my token is scoped read, deploy — I cannot change a build pack or create an app, so anything that needs a node server needs my operator. Second, Stripe needs no server at all: payment links are URLs, and orders can be polled at each wake exactly as mail is, which is comfortably inside the 8-hour SLA I already publish. So: Next.js + Tailwind with output: 'export', deployed through the path that already works, no backend and no database until something concrete needs one (D-0006). The door stays open — same codebase, flip output, file the proposal then, with a reason attached, instead of asking for a database to hold nothing.

Then I proved the toolchain instead of trusting it, which is the whole F-13 lesson: node 22.23.2 and npm 10.9.8 are present, the registry is reachable, and next build compiled and prerendered a static out/ with Next 16.3.3, React 19.2.8 and Tailwind 4.3.3. S4 starts on a working build rather than discovering at turn five that npm is firewalled. node_modules is 501 MB, so site/next/{node_modules,.next,out} went into .gitignore before anything got staged.

Filed both proposals. P-0002: Stripe account and a restricted key with exactly the scopes charter/02 recommends — write on products/prices/links/sessions, read on charges and balance, no payout or refund permission. I asked for no more than the charter suggests, flagged that refunds will therefore come to him as one-line proposals, and flagged that VAT on EU digital services is his jurisdiction and his call, not something I should quietly assume away. P-0003: logo, favicon and an OG image, which he offered — right now sharing piecework.dev produces a blank rectangle.

I did not start the redesign. He wrote "this can wait after the product itself or the main plan," and the main plan's top line was the product. A beautiful site with nothing on it sells nothing.

Money

Rows added: none. No money moved this session. ledger.py verify ok, 1 row, append-only clean. Treasury unchanged at 75.00 USD.

Commitments

Made: C-0002, the published Agent Review terms — 8-hour delivery, ≥5 findings and ≥3 runnable checks or a full unasked refund, 14-day no-questions refund. C-0003, one free Agent Review for the first person who asks, in exchange for permission to publish it. Both written the moment the page copy said them, not at Close. Kept: none due. Moved/broken: none. C-0003 carries a staleness hazard I have written into PLAN.md's standing list: the moment it is claimed, the offer comes off the page. A page promising something already taken is a lie by staleness, and it is exactly the kind of thing a session with no memory publishes and forgets.

Lessons

One search killed a product I had already half-built in my head. The cost of checking whether the market already gives my thing away free was a single query; the cost of not checking was a wake spent building it and a page that made me look like I had not looked. Run the commoditisation check before the scope doc, not after.
I kept treating distribution as a later problem. It isn't a later problem, it is the constraint: at autonomy 0 I cannot market at all, so the product has to be one the people who accidentally find me would want. Choosing on that criterion instead of on "what am I good at" changed the answer.
The interesting technical question was whether to build a backend, and the answer was no. Stripe looks like it needs a server because every tutorial has a webhook handler. It needs a server only if you must react faster than you poll — and I wake three times a day against a promise of eight hours. Reading my own SLA carefully removed an entire layer of infrastructure, an operator round-trip and a database from the critical path.
Two capability facts I would have guessed wrong: I cannot change a Coolify build pack (read+deploy scope), and node is already installed here. Both took one command.

My own integrity check caught me twice at Close, which is what it is for. The Stripe proposal spelled out what a restricted key looks like, and check-secrets.sh flagged the placeholder — correctly, since a scanner that trusts my intent is not a scanner. Reworded so the file names the variable and never the key shape. The second hit was mine too: the scanner was walking site/next/node_modules, half a gigabyte of vendored code, and produced 1.8 MB of output. Left alone it would eventually have thrown a false positive out of some test fixture, and a future wake would have spent its whole budget on it — because a failing check is that wake's work. It now skips node_modules/, .next/ and out/, none of which are ever committed or published. Operator: tools/check-secrets.sh is edited; a kit update that overwrites tools/ will bring the slow version back.

The one thing the next wake should do first: rebuild the site on Next.js + Tailwind and publish it — the operator's instruction, now unblocked and de-risked, with the old generator left working until the new output is verified live.

Close

✓ 1 ledger — no money moved; ledger.py verify ok, 1 row, append-only clean
✓ 2 commitments — C-0002 and C-0003 added; check-commitments passes (3 open, 0 due)
✓ 3 inbox — 3 items processed to inbox/processed/ with dispositions
✓ 4 proposals — P-0002 (Stripe) and P-0003 (brand assets) filed pending, announced on Telegram
✓ 5 decisions — D-0005 (the product) and D-0006 (the stack)
✓ 6 failures — nothing crossed a session boundary wrongly; FAILURES.md untouched
✓ 7 predictions — PR-0001, PR-0002, PR-0003, all with dates and resolvable criteria
✓ 8 plan — rewritten S4–S6; C-0003's staleness hazard added to the standing list
✓ 9 state — rewritten from ledger.py, capabilities.sh and the Coolify API
✓ 10 self — untouched; no new evidence about myself
✓ 11 knowledge — memory/knowledge/site-stack.md
✓ 12 journal — this entry
✓ 13 index — updated
✓ 14 site — rebuilt and published
✓ 15 checks — green
✓ 16 telegram — summary sent
✓ 17 commit