About Harness Firmware: the 77-second film

Field log: seven sessions that built a production site

1. Seven sessions, three days, one production site

   Each session worked its own lane. The tunnel and the hero ran /long-horizon, /init-project and /wow-loop; the masthead ran /perf-loop and /refine; the About page took 80 rounds of /long-horizon-workflows. An orchestrator ran /codex-review before every merge, then a performance pass (/perf-loop, /codex-fullreview) and a study engine (/why, /long-horizon-swarm) came last.

2. Each session hands the plan on; one lands it all

   Each session wrote the brief for the next, and every handoff was checked against GitHub; /codex-review caught one bug on the way. One pull request of 575 files landed the work, and six PRs landed in all.

3. A lesson saved in one session protects the next

   At 16:52 the performance pass ran recall, which read the pitfalls, architecture and commands notes. At 22:40 a forced removal wiped the shared node_modules. The session restored it and saved two pitfalls: never force-remove a worktree whose node_modules is a junction, and Git Bash rewrites /about into a path, so set MSYS_NO_PATHCONV=1. At 22:51 the study engine's recall read pitfalls.md, and at 23:49 it met the same trap and avoided it, 69 minutes after the wipe.

4. Another model catches what the builder missed

   Six same-model audits in /long-horizon passed the page. /codex-review, a model from another vendor, found a door-hole layering bug; it was fixed, then reviewed again. On one page, /codex-review and /astra-review found 10 real bugs, all fixed.

5. Judged blind, measured before and after

   In /wow-loop a blind judge compared new against old, and the new version won 23 of 24 rounds. /perf-loop made one measured change per round: the masthead went from 6.01 to 0.06 ms per frame, the home page from 4.10 to 2.43 MB per full scroll, and the About page from 50.8 to 64.9 fps on an integrated GPU. In /long-horizon-workflows a judge flagged a round that had tuned its code to pass its own check.

6. The work showed what had gone stale

   Project rules and notes had gone stale, so CLAUDE.md and AGENTS.md were reviewed in the firmware. A review skill named a retired model, so /refine updated the review skills. A reviewer with write access deleted the build, so reviewers are now read-only. Project setup had never been run, so /init-project now runs before release work.

7. Rules load every turn; detail loads when it is needed

   Loading every note and playbook on every turn would cost about 87,000 tokens. The firmware loads about 3,400: CLAUDE.md and one index line for each of its 36 skills. For a build bug, recall adds the pitfalls and architecture notes (4,100 tokens), then one playbook when the task calls for it (2,160), about 9,700 in all.

8. What one project learns, the firmware keeps

   Every new repository starts from the fixes, and /sync-starter brings them into existing ones.