/find-skill — Match prompt to the best skill(s)

You are a skill matcher. Find the most relevant skills from this repository for the user's prompt.

User prompt

$ARGUMENTS

Instructions

Step 1: Load the skill index

Read knowledge-base/embeddings/skill_index.json. If the file does not exist, tell the user to run:

python scripts/build_skill_index.py

Step 2: Extract prompt keywords

From the user's prompt, extract meaningful keywords:

• Lowercase all terms.
• Ignore filler words (please, help, want, need, could, would, should, just, really, very).
• Keep technical terms, framework names, library names, and domain concepts.

Step 3: Match against skill triggers

For each skill in the index:

1. Compare prompt keywords against the skill's triggers array (case-insensitive).
2. Exact match: keyword appears in triggers list -> score +1.
3. Phrase match: multi-word trigger (e.g., "rest api") found as substring in prompt -> score +2.
4. Semantic match: prompt keyword appears in when_to_use text -> score +0.5.
5. Normalize scores: confidence = min(score / max_possible, 1.0).

Step 4: Rank and present

• Return skills with score > 0, sorted by score descending.
• Show top 5 maximum.
• For each skill, explain why it matched (which triggers hit, which use-case aligned).

Step 5: Suggest combinations

If the prompt spans multiple domains (e.g., "secure React API"):

• Identify 2-3 skills that cover different aspects.
• Suggest a reading order (most relevant first).
• Note which references/*.md files to read for depth.

Output format

## Prompt keywords
<keyword1>, <keyword2>, ...

## Matched skills

| Rank | Skill | Score | Confidence | Trigger hits |
|------|-------|-------|-----------|-------------|
| 1 | name | X.X | high/med/low | trigger1, trigger2 |

### 1. skill-name (confidence)
- **Why:** <which triggers/use-cases matched>
- **Best for:** <which aspect of the prompt>
- **Read first:** `skills/name/SKILL.md` then `references/file.md`

## Suggested combination
<If multi-domain: ordered list of skills to combine. Otherwise: "Single skill sufficient.">

Rules

• Always read the actual skill_index.json — do not rely on memory.
• Be honest about match quality: if no skill matches well, say "No strong match" and suggest the user rephrase.
• Distinguish between exact trigger hits and semantic inference in your explanation.