Learning from agent failures
Skill casioreview20-glitch/forge-os/skills-v2/kernel/learning-from-agent-failures
Use when repeated agent, route, tool, context, or verification failures reveal a reusable technique gap rather than a one-off project defect.From its SKILL.md
npx -y skills add casioreview20-glitch/forge-os --skill learning-from-agent-failuresAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 20 days oldThe repository was created 20 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 11 stars11 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its file declares
Copied from the file, not written here
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
0.8 KB, 95 tokens by cl100k_base, as published. Nobody here has run it
Learning from Agent Failures
Core principle
Cluster failures by invariant, rationalization, missing context, and tool boundary. Propose a skill or code change without auto-promoting it. The runtime owns deterministic scope, coverage, policy, and evidence checks; the agent owns only the judgment that cannot be reduced safely to code.
Do not activate when
- the failure is caused by a known code bug
- one isolated failure has no reproducible pattern
What ships with it: 9 files
11.4 KB alongside SKILL.md
evaluators/
- baseline.json453 B
- cases.json651 B
- rubric.json416 B
sections/
- decision-tables.md391 B
- examples.md355 B
- failure-modes.md218 B
- procedure.md549 B
- verification.md241 B
- manifest.json8.2 KB