Resolved errors should stay resolved

The loop that turns a failure in one session into a rule for every session after.

From failure to standing rule

A failure caught in one session is logged by the checkpoint skill and turned into a standing rule in the agents' instructions.

Three of the real rules are shown here. Every later session reads them before it starts, so the same mistake doesn't come back.

At Formative Minds most of the code and media was written by AI agents in Claude Code, orchestrated through custom skills. Each agent session starts fresh, with no memory of the one before it.

problem

Because every session starts from zero, a mistake fixed in one session could come straight back in the next. Correcting the agent again each time doesn't scale, and it relies on me remembering every past failure.

solution

Every checkpoint writes a devlog entry: what shipped, why, and what went wrong. Failures from the devlog are turned into standing rules in the agents' instructions, which every later session reads before it starts.

Some examples: a shortened ID once silently loaded the wrong asset, so the rule is to always pass the full ID. Agents reported UI as finished without being able to see it, so visual claims are now tagged as not visually verified. The same fact typed into several files drifted apart, so every fact now has one home, with a test that keeps any copies in sync.

The generative media playbook alone holds 28 of these rules.