A game jam where no broken fix is tried twice

Solo buildersHobby game developer doing a 48-hour jam in Godot

We entered a 48-hour jam alone, working in bursts between naps. After each sleep, Claude Code had forgotten the theme rules, the web export limits, and that the jump physics had already been rewritten twice.

How do I stop Claude Code redoing failed fixes and forgetting my jam rules after every break?

With ArcRoom, a failed approach is saved with remember_failure and what worked instead, so the next session is warned before it tries it again. Save the jam rules with remember at the start: theme, must run in the browser, no outside assets. A done-check in .arcroom.json can run godot --headless --quit to catch script errors before the agent says done, and the last-run note says what the last session was doing.

Also asked as: Claude Code forgets my Godot project after every nap. How do I fix that? · How do I stop my AI agent from rewriting the same jump physics again? · Can an AI agent remember game jam rules across sessions?

The work

At the start, remember the jam rules: theme, 'must run in the browser', 'no outside assets'.

ArcRoom keeps
  • Last-run note
  • Done-checks
  • Saved decision
The next agent gets

last session (4 h ago) stopped mid-way.

Step by step

  1. At the start, remember the jam rules: theme, 'must run in the browser', 'no outside assets'.
  2. After each nap, the last-run note says what the last session was doing and whether it stopped mid-way.
  3. A failed physics approach goes in with remember_failure so it isn't tried a third time.
  4. A done-check in .arcroom.json runs godot --headless --quit, which catches script errors before the agent says done.

What the agent sees

ArcRoom: last session (4 h ago) stopped mid-way.
Asked: "add the second level and the exit door"
Tried and failed: CharacterBody2D jump with custom gravity. Why: jitter on slopes.

Compared with today's tools

A TODO.md the agent updates

Ask the agent to keep a running TODO and notes file, and point each new session at it.

Setup
5 min
Ongoing
remind the agent to update it before every break, ~2 min

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min (npx arcroom init)
Ongoing
None

Works if the agent remembers to write it before you fall asleep; ArcRoom records the steps either way and blocks the repeat dead end.

Where it stops

A jam is mostly new work, so there is little to reuse. The value is in restarts, not repeat answers.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in solo builders