Tune an instruction a workflow will reuse

Anchoring an ambiguous output shape to concrete examples, then scoring competing phrasings against real test cases -- prompt tuning with evidence, not just a rewrite by feel.

Also usesclaude-plugins-official

Anchor an ambiguous output shape to concrete examples before reaching for more descriptive prose, then score competing phrasings against real test cases rather than picking by feel. If the tuned instruction is destined for a SKILL.md, its frontmatter and description are what decide whether it fires at all, so settle the wording before shipping it there.

Start here

One prompt to start

This reused instruction's output keeps coming out wrong -- anchor the shape to a few concrete examples instead of more descriptive prose, then score competing phrasings against real test cases rather than picking by feel, check the frontmatter rules if it's going to live in a SKILL.md, and finally run an effort-ladder check on the settled wording's own failure risk.

Needscompass, plugin-dev

Beats

Run these in order. Each prompt is copy-pasteable straight into Claude Code.

1. compass:compass-calibrate-format

   Anchors an ambiguous output shape to 2-5 concrete examples instead of adding more prose describing the shape.

   Run this beat on its own

   this instruction's output format keeps coming out wrong -- give me a few concrete examples of exactly the shape I want instead of more description

2. compass:compass-optimize-instruction

   Generates one candidate per framing and scores each against real test cases -- explicitly not intended for a one-off prompt with no test cases to score against.

3. plugin-dev:skill-developmentalso in 1 other recipe

   • Author a new skill or agent for a plugin

   If the tuned instruction is going to live inside a SKILL.md, the frontmatter and description rules are what actually decide whether it ever fires.

4. compass:compass-reason-verify

   A 4-rung effort ladder applied to the resulting instruction's own failure risk, once the wording itself is settled.

Worked example

Grounded in — why this beat order is trustworthy

tools/plugin-serializer/'s generator prompt: this repo's CLAUDE.md records that its YAML-frontmatter traps had to be named explicitly in the prompt before sonnet reliably stopped producing them, and that opus bought nothing extra on the same task -- a tuned instruction with a measured pass/fail history, exactly beat 2's precondition.

Do / Don't

• ✓Anchor an ambiguous output shape to 2-5 concrete examples before adding more descriptive prose.
• ✓Score competing phrasings against real test cases -- this only works when real test cases actually exist to score against.
• ✓Check the frontmatter and description rules if the tuned instruction is destined for a SKILL.md -- they decide whether it fires at all.
• ✓Run the effort-ladder check on the resulting instruction's own failure risk only once the wording itself is settled.

• ×Don't add more descriptive prose to fix an ambiguous output shape -- anchor it to concrete examples instead.
• ×Don't reach for compass-optimize-instruction on a one-off prompt with no test cases to score against.
• ×Don't ship a tuned instruction into a SKILL.md without checking whether its frontmatter and description will actually make it fire.