The Playbook

The habits, with the evidence

1. 01

   Enforce rules in code, not only in prompts

   • CityBrew

     Kill switches that default to paused, and a hard failure instead of ever emailing an empty digest.

     Source: .github/workflows/weekly-digest.yml

   • Reviresco

     Concepts are screened twice for trademark risk: once by the model, then by a denylist in code. Killed ideas are kept with the reason, never dropped.

     Source: revirescosupplyco src/lib/screening.ts

   • 3D printing

     After three jams, Claude read the printer's own logs and found the signal: a steady 15% drop in heater power at the same temperature means plastic has stopped flowing. A watcher script checks for it, and every print file must pass a preflight check first.

     Source: 3D Printing project notes, jam investigation, October 2

   • Side quests

     Twenty tools for an AI agent to manage a fantasy team. Every write plans, refuses if blocked, runs dry by default, executes once on confirm, verifies, and logs exactly one audit row.

     Source: ClaudeCode sleeper-agent/README.md

   • SignalWall

     Every release checks itself in its own environment, then must prove it is healthy after restart or it flips back.

     Source: docs/SPEC.md, Auto-update

2. 02

   I make the judgment calls

   • Reviresco

     The cinema went from design boards to working pages in about fifteen minutes, then gained Wallet tickets and a seat picker.

     Source: goreviresco cinema, 2026-09-10

   • Three small Mac apps

     The fan icon shipped spinning at 30 frames a second in code and once a second on screen. The fix handed the spin to the system compositor.

     Source: BetterBattery ca02323

   • 3D printing

     The first replacement was my idea: drop the rod into a cradle and hold it with a magnet. It didn't work, so I scrapped it and asked for a snap claw like the stock one, only stronger. That's the clip in the deep dive.

     Source: Printer setup chat, September 30 and October 1

3. 03

   Write the spec before the code

   • Home Manager

     Product brief, requirements, architecture and UX design, each built step by step with a planning method for AI agents, all on the first day.

     Source: home-manager-app _bmad-output/planning-artifacts

   • SignalWall

     No canvas size is hardcoded: every golden frame renders at 128 and 192 dots wide.

     Source: CLAUDE.md, Non-negotiables

4. 04

   Automate manual steps

   • Three small Mac apps

     An in-place updater that installs whatever app the release contains, so renaming the app could not break it.

     Source: BetterBattery Sources, UpdateChecker.swift

5. 05

   Break work into session-sized tasks

   • CityBrew

     An autonomous build loop shipped five epics of stories in about 80 minutes; the evening went to human testing and fixes.

     Source: CityBrew history, 2026-04-28, 13:05 to 14:24

6. 06

   Design for cost

   • CityBrew

     A cost briefing found most of the spend was web search during a daily scrape and a server running all day to fire a timer. The redesign made it one short weekly run.

     Source: docs/COST-AND-ARCHITECTURE-BRIEFING.md

7. 07

   Keep project state in the repo

   • Three small Mac apps

     Rules that came out of bugs, like never animating with a timeline view in a menu bar panel, are written into each repo's instructions so the next session starts knowing them.

     Source: BetterBattery, DayDial, NetPulse CLAUDE.md

8. 08

   Record decisions and how to reverse them

   • SignalWall

     45 architecture decisions on record, each with an escape hatch. A decision changes only by a new entry that supersedes it.

     Source: SignalWall docs/DECISIONS.md

9. 09

   Retire a build when something simpler works

   • CityBrew

     Retired once a Claude scheduled task and a skill could do the same weekly job with no servers, no database and no deploys.

     Source: Owner, Workshop kickoff, 2026-10-06

10. 10

    Run agents in parallel on separate files

    • SignalWall

      Orchestrating agents ran builders in separate worktrees, each owning its own files, working through work orders sized for one focused session.

      Source: SignalWall STATE.md; docs/build/S8.md

11. 11

    Test against recorded API responses

    • SignalWall

      Every data source is tested against a recorded real response. No test ever calls the live service, and no API shape is written from memory.

      Source: SignalWall docs/CONVENTIONS.md, Tests