| name | Caddy Explain |
| description | When a [CADDY] hint fires and the principal is confused why, /caddy-explain walks the scoring transparently — which trigger matched, what points it earned, what the threshold was, what other skills were close. Pairs with the legibility thesis. Caddy's decisions should be inspectable on demand. |
Caddy Explain
When Caddy surfaces a hint and you don't see why, run /caddy-explain to see the deterministic scoring. The scoring should never be a black box.
When to use
- A [CADDY] hint fired and you're surprised by the choice
- A [CADDY] hint didn't fire when you expected it to
- You want to verify why a particular skill won (or lost) for a recent prompt
- Triggered by Caddy on prompts like: "/caddy-explain", "why did Caddy suggest", "explain that hint", "why no hint", "why this skill", "Caddy scoring", "Caddy why"
How to execute
Step 1 — Identify the prompt being explained
If the principal provides the prompt explicitly, use it. Otherwise:
- Use the immediately-preceding user prompt from the current session
- If there's ambiguity ("the one before the last hint"), ask one focused question
Step 2 — Walk the scoring deterministically
For each skill in .claude/skill-index.json:
- List its triggers
- For each trigger, check whether it appears in the prompt:
- Multi-word phrase match (case-insensitive substring) → 3 points
- Single-word boundary match (
\bword\b regex) → 1 point
- No match → 0 points
- Sum the per-skill score
Step 3 — Show the result
Output format:
🔍 Caddy scoring for prompt:
"{prompt text}"
Threshold: 2 (skills below this score are silent)
Max hints: 2 (top scorers above threshold are surfaced)
SURFACED:
/skill-a — score 5 — matched: "phrase match" (3pts), "keyword" (1pt), "another" (1pt)
/skill-b — score 3 — matched: "exact phrase" (3pts)
NEAR MISSES (below threshold):
/skill-c — score 1 — matched: "word" (1pt)
/skill-d — score 1 — matched: "another" (1pt)
SILENT (no triggers matched):
/skill-e, /skill-f, /skill-g, /skill-h, ... (28 more)
Reasoning:
Caddy chose /skill-a + /skill-b because their scores cleared the threshold
of 2 AND they were the top 2 (max 2 hints per prompt).
/skill-c and /skill-d had matches but didn't clear threshold.
Step 4 — If the principal expected a different result
Ask: would they like to:
- Adjust triggers for a skill that should have matched but didn't (one trigger could be added)
- Narrow triggers for a skill that fired when it shouldn't have (a trigger could be removed)
- Accept the result and move on (sometimes the scoring is correct and the surprise was misremembered)
If they pick adjust/narrow, they can run /caddy-enroll <skill-path> to update the index, or edit .claude/skill-index.json directly.
Output principles
- Show the math. Don't summarize "the skill matched well." Show the trigger, the points, the threshold.
- Show the near-misses. The skills that almost matched are often the most informative for tuning.
- Don't apologize. If the matcher chose correctly, say so plainly. Caddy is deterministic; surprise often means the principal's mental model needs updating, not the matcher's.
- Don't auto-tune. This skill explains; it doesn't change the index. Auto-tuning is a separate decision the principal makes after seeing the explanation.
Why this skill exists
The legibility thesis says every layer of the framework must be inspectable. Caddy is one of the most-fired layers (every prompt). Without /caddy-explain, its decisions become a black box — which violates the thesis and erodes trust over time.
/caddy-explain makes Caddy's decisions auditable on demand. The scoring is deterministic; the explanation should match.
Connects to
- [[Caddy]] — the doctrine note this skill makes inspectable
- [[Suggestion Credibility]] — the constraint Caddy operates under, including inspectability
- [[Legibility Thesis]] — the meta-doctrine this skill enforces at the Caddy layer
