skillbuilder

/skillbuilder

Build Claude Code skills that are genuinely intelligent — not just format-valid, but skills that exploit something specific about how Claude works.

Reference standard: the skills in this repo. Study them before drafting.

Skill	What it exploits
/stochastic	LLM variance — poll N runs, aggregate consensus/divergences/outliers
/model-chat	Disagreement — multi-agent debate surfaces assumptions a single pass misses
/fanout	Parallelism — independent researchers cover more ground, Opus synthesizes
/autoresearch	Measurement loops — read code, mutate, measure, keep-or-revert, repeat

Every good skill exploits something. The question isn't "which of 8 buckets" — it's "what does this skill do that a raw prompt can't?"

Quality bar

A skill is worth building if:

1. Reusable — works across many tasks, not a one-shot
2. Exploits something — LLM variance, parallelism, measurement, memory, gates, whatever
3. Beats a prompt — materially changes Claude's execution, not just vocabulary
4. Passes the removal test — delete the skill, output measurably degrades

If the answer to "what does this exploit" is "nothing, it's just instructions" — that's a prompt, not a skill.

Anti-patterns

1. Task helper. "Skill that writes commit messages." That's a prompt. "Skill that spawns 3 agents to review, vote, kill weak verbs" — that's a skill.
2. Best-practice list. Documentation isn't a skill. A checklist that gates SHIP/NOT SHIP is.
3. Prompt-equivalent. If a one-shot prompt produces the same output, don't build a skill.
4. Single-use. "Skill to set up my repo." That's a script.
5. Decoration. Skill exists to feel useful but doesn't change execution.
6. Token bloat. SKILL.md > 250 lines of best-practice advice Claude would do anyway.
7. Generic naming. /helper, /assistant, /improve-X. Names should name what it does or what it exploits.

Operations

craft — build a new skill

Input: domain or problem. Output: working skills/{name}/SKILL.md.

Flow:

1. Study references. Read existing skills in this repo. What patterns apply?
2. Identify the exploit. What will this skill do that a raw prompt can't? If nothing — stop, recommend a prompt instead.
3. Draft the skill following the template below.
4. Write to disk. skills/{name}/SKILL.md
5. Report. Path, what it exploits, restart required.

audit — check an existing skill

Input: skill name or path. Output: lands · needs work · kill it.

Flow:

1. Read the SKILL.md.
2. Identify what it exploits. Implicit or missing → needs work.
3. Run anti-patterns 1-7. Any hit → needs work or kill.
4. Removal test: would Claude's output degrade without this skill?
5. Verdict + one move.

kill — diagnose a failing skill

Input: skill name + symptom. Output: fix or kill.

Flow:

1. What was it supposed to exploit? Has it drifted?
2. Anti-pattern check.
3. One move: sharpen, narrow, or kill.

Skill template

---
name: {kebab-case}
description: "TRIGGER when {specific signal}. {What it does in one phrase}. SKIP for {anti-domain}."
user-invocable: true
allowed-tools:
  - {minimum needed}
effort: {default | high}
---

# /{name}

{What this skill does. What it exploits — why it beats a raw prompt.}

## Reference

{If applicable: similar skills, shipped products, or patterns this builds on.}

## How it works

{The mechanism. What Claude does differently when this skill is loaded.}

## Anti-patterns

{Domain-specific failure modes to avoid.}

## Output

{What the skill produces. Format, structure, deliverable.}

Frontmatter spec

Field	Required	Notes
name	yes	kebab-case, matches folder name
description	yes	TRIGGER when... SKIP for... — this is how Claude decides to load it
user-invocable	yes	true if user can call it directly
allowed-tools	yes	minimum needed — don't over-grant
effort	yes	default (single-pass) or high (spawns agents, long-running)

Allowed-tools — minimum needed

• Read — almost always
• Agent — if it spawns sub-agents
• Bash Glob Grep — for codebase work
• Write — only if it persists state
• WebSearch WebFetch — only if it does live research

Don't grant tools the skill doesn't use.

Naming

• Mechanism-named — /stochastic, /fanout (names the how)
• Domain-named — /autoresearch, /skillbuilder (names the what)
• Avoid — /helper, /assistant, /better-X, /improve-Y

After writing

1. Tell user to restart Claude Code
2. Test: invoke in fresh conversation, verify trigger matches
3. Check: does output beat a raw prompt?

Voice

• Lead with what it exploits
• Every claim cites a reference (existing skill or shipped product)
• If it's a prompt in disguise, say so
• Short sentences, specific mechanisms
• One skill per run