Make the spec fail the build: a solo dev's log of docs-versus-code drift

A developer who describes himself as a coffee roaster in Taiwan, eight months into teaching himself to code with AI, has published the prompt log of the day he stopped asking his assistant to fix bugs and started asking it to block them ^[1]. By his own account the shift was from "fix this" to "fix this and prevent it from happening again" ^[18], and the thing that came out of it was a CI script that blocks pushes on rule violations ^[6]. That is the right shape, and it is worth looking at closely because the log also shows how far the shape still is from the promise.

The timestamps are unusually precise. At 11:41 on April 21, 2026 he asked whether three audit rules should go into something called governance-guard, then be used for a targeted scan ^[3]. At 15:08 the same day he chose the CI option: scan all relative and alias import paths, fail equals block ^[4]. At 01:12 the next morning he asked whether the fixes and rules could be saved as a reusable "skill", because he expected the work to be forgotten once his attention returned to the MVP launch and to turn into hidden risk ^[5]. That is thirteen and a half hours from idea to enforcement policy ^[15], inside one session, which produced governance-guard.mjs, a TECH-DEBT-BACKLOG.md where known violations get queued rather than fixed, and skill files meant to survive across projects ^[6].

The motivating defect is the interesting part. On April 29 he wrote that a previous balance was still showing zero after five or six versions of fixes ^[8]. The cause, as he describes it: the docs said single_price was the single source of truth, the code used a five-layer fallback, and the UI was still recalculating, so fixing one site left four others in place ^[9]. It became one of fourteen data-drift debugging records and produced a written instruction, Rule 17 Step 0, that value-mismatch bugs must begin with a SQL query against the database rather than a guess in code ^[10][11]. April was his heaviest month at 348 commits, the project's single-month high ^[7], roughly 11.6 a day ^[16].

So the diagnosis is a docs-versus-code divergence, but the executable check that shipped is an import-path linter ^[4]. The account does not describe an automated test that compares the written spec to the implementation, and publishes no before-and-after defect numbers ^[17]. The SSOT rule stayed prose; the enforcement went to a different, cheaper rule. That gap is the honest state of this idea generally. Import paths are trivial to assert in CI. "The code has exactly one price authority" requires the spec to name the authority in a machine-readable way, and the check to fail when a second path appears. Nothing here says that was built.

Two things to watch. First, whether the skill files actually carry rules into a second project, which is the claim that separates a governance primitive from one repo's lint config ^[5]. Second, whether Rule 17 Step 0 ever becomes a test ^[11], or whether the backlog file ^[6] simply absorbs violations at the rate a solo developer can triage them.