What's this about?
Warp, the terminal company, tried using an agent for internal code review and it did not work well. The diagnosis: the agent didn't know their project, didn't know the team's conventions, didn't know the history of what had gone wrong before — and when told about a problem, didn't remember next time. No memory. They first tried the obvious patches: hand-editing the system prompt from failure cases, and filling out AGENTS.md. Neither worked well enough. The post describes what they built instead — a skill that evolves from its own failures.
What is the skill-evolution problem?
A skill is a packaged, reusable set of instructions. Static skills go stale and never learn from mistakes. Self-updating skills learn — but they grow without bound, and if the domain has no objective standard, they get *worse*: dotey reports his self-evolving JS-decompilation skill works well and keeps improving while the file grows ever larger, while his self-evolving writing skill was "hard to describe", frequently negatively optimising, because writing has no fixed criterion to improve against.
What's the community saying?
- The prerequisite the post makes clear: self-improvement needs a verifiable signal. Code review has one — did the bug ship? Writing does not.
- dotey's aside that people enjoyed: the post is published by Anthropic but uses AGENTS.md rather than CLAUDE.md.
- It is the industrial counterpart to WikiSkill and JIT-Agent from earlier in the week — three independent arguments that durable value sits in the scaffolding.