claude-code - 💡(How to fix) Fix Skills should self-recommend effort level after execution [1 participants]

Problem

Skill manifests support an effort frontmatter field (low, medium, high, max), but there's no guidance on what to set it to. The only way to tune it is manual trial and error — run the skill, eyeball the output, decide if it over-thought or under-thought, edit the YAML, repeat.

Claude has all the signal needed to make this judgment itself:

• Whether the task was mechanical (sequential tool calls, no branching) or required synthesis
• Whether it had to backtrack or reconsider
• Whether the output quality would have degraded at a lower effort level
• Token usage relative to task complexity

None of this feeds back into anything.

Proposal

After a skill executes, Claude should be able to emit an effort recommendation, e.g.:

"This skill ran at high effort but the task was mechanical — low would produce equivalent results and run faster."

This could take several forms (not mutually exclusive):

1. Post-execution advisory — Claude notes in its response when the effort level seems mismatched
2. --recommend-effort flag — Run a skill once and get a suggested effort level for the frontmatter
3. Auto-calibration — After N runs, suggest an effort level based on observed reasoning patterns

This is low-hanging fruit — the model already knows whether it needed to think hard. It just doesn't say so.

Why this matters

Skill authors (especially in orgs with many shared skills) currently have to guess. Too high wastes tokens and time on every invocation. Too low produces shallow results. The model is the only one who actually knows which it needed.