Skill Feedback Cycle
An iterative quality loop for improving Claude Code skills: invoke → watch → give feedback → skill updates itself → repeat. Step 6 of the Six-Step Skill Building Framework, and the mechanism by which a skill that initially produces generic output converges on production-grade output over 10–30 runs.
Key points
- Each cycle: invoke the skill, watch it work, identify what to correct, give feedback, skill updates its own SKILL.md [src-013]
- “The first couple times you run a skill, you may feel like it’s very AI generated. But by the time you’ve run that skill 10, 20, 30 times, every single time it gets better.” [src-013]
- The skill itself should contain the feedback-loop instruction — asking the human for a quality score and patch notes at end of each run [src-013]
- This is an application of the Curiosity Rule: never accept output passively; always interrogate why it made the choices it did [src-013]
- In Hermes, Nate frames skills as the “how to do it again” half of the assistant. If the user corrects Hermes on the same workflow repeatedly, the correction should become a skill or a skill patch rather than another one-off prompt [src-074].
- Hermes can create, update, and discover skills through the skills hub/community pattern, but Nate still treats human review as necessary because trigger frontmatter and procedural detail determine whether the right skill fires [src-074].
Related concepts
- Six-Step Skill Building Framework — parent framework; this is step 6
- Curiosity Rule — mindset prerequisite for using this cycle effectively
- AIOS Daily Loop — the daily cadence that surfaces improvement opportunities
- Progressive Context Loading (Skills) — what the skill file contains that gets updated each cycle
- Hermes Five-Pillar Agent Architecture
- Agent Harness Portability