find-me-a-skill

find-me-a-skill

SKILL.md is the prompt Claude follows; this file is for you.

Works on its own.

What it does

Tells you which skill or sample workflow fits a task, which model and /effort level to run each step at, where you'll be asked to decide something, and the exact commands to paste. It suggests and stops: it never starts the work itself.

When to use

• "Which skill should I use for this", "find me a skill", "how should I run this", "what model / effort for this".
• At the start of a session, with no task in mind: it reads the top board tasks and suggests one option for each.
• Not for running anything: paste its commands to do that. Not for a list of every skill either; that's the site's home page or the CATALOG.md it reads.

How to invoke

/find-me-a-skill <task>
/find-me-a-skill

Model-invocable, so the trigger phrases above work too. With no text it uses the conversation's goal, then the board.

How it works

1. Task.
   • The text after the command.
   • Else the goal of the conversation so far.
   • Else the top 1-3 tasks on the repo's board, found the way next does (its BOARD.md).
   • No repo or no board: it asks what the task is.
2. Candidates. Every skill Claude can see in this session (this collection, plugins, built-ins) plus the entries and sample workflows in CATALOG.md.
3. Options, at most 3.
   • The primary pick is the best fit from everything installed.
   • When an entry or workflow from this collection fits and isn't already the pick, it adds that too.
   • Any option from the catalog is labelled "from your collection".
   • A catalog option is labelled "not installed", with the install command, when it isn't in this session's skill list, in ~/.claude/skills/<name>/ or in the plugin's install dir. With no plugin installed, that last check is skipped and the reply says so. User-only skills never show in the session's list, installed or not.
   • Plugin installs get the krish: prefix on commands.
   • Nothing fits: it says so, names the nearest option and suggests doing the task directly.
4. Model and effort per step.
   • Starts from CATALOG.md's defaults. For a large or risky task, effort goes up first, and the model only once effort is at max. A small task works the same way in reverse.
   • A skill that isn't in the catalog shows its origin (plugin or built-in) and an estimate label.
5. Reply. Per option: why it fits, the steps with model and effort, the points where you decide, then /model, /effort and the invocation for option 1, ready to paste.

Reads: CATALOG.md, this conversation, the board (via gh) in board mode. Writes: nothing.

Design principles

• Suggest, don't run. It runs on Haiku at low effort, so it must not start heavy work by accident; you start it by pasting the commands.
• One generated catalog. CATALOG.md is generated from the repo (python scripts/site_catalog.py), and a test fails when it drifts. So the skill works offline in any repo, and its defaults match the site's "suggested model · effort" chips.
• Broad first, own collection second. The best tool for the task wins even when it's a plugin or built-in; this collection gets a second slot only when it genuinely fits.

Use cases

• You know the task but not which skill or sample workflow covers it.
• Picking a cheaper model or lower effort for a routine step, and a stronger one for the step that needs judgment.
• Starting a session cold: "what should I do next, and how should I run it".

Tips

• The model: haiku and effort: low frontmatter applies for this turn only. If your setup ignores the override, run /model haiku first, then switch to the suggested model before pasting the commands.
• Say the size or risk ("big refactor", "one-line fix") and the suggestion moves a tier or level accordingly.
• Defaults live in site/data/meta.json (model_effort). After editing them, or any skill description or workflow, regenerate the catalog with python scripts/site_catalog.py.

Example

Three runs, each traced to entries in CATALOG.md: a task in words, an idea to test, and an empty prompt that falls back to the board.

> /find-me-a-skill address the review comments on PR 12
⏺ Read(CATALOG.md)
  ⎿ | pr-watch | skill | sonnet | high | `/pr-watch` | Watch the user's open PRs to merge ... | pair-to-pr |
⏺ Task: address the review comments on PR 12
  1. pr-watch (skill, from your collection): reads every review thread, fixes the valid ones via pair, replies to or resolves the rest with evidence.
     Steps: /pr-watch (sonnet · high) -> pair -> manager agent (opus · high) -> implementer agent (sonnet · medium)
     You decide: each push, reply and resolve; a force push needs its own yes.
  Run option 1:
  /model sonnet
  /effort high
  /pr-watch 12
> /find-me-a-skill I want to stress-test an idea for a plugin system before anyone builds it
⏺ Read(CATALOG.md)
  ⎿ ### pitch-an-idea: Pitch an idea against the plan
⏺ Task: stress-test a plugin-system idea against the plan
  1. pitch-an-idea (workflow, from your collection): grills the idea one round at a time, then checks it against open issues and recorded decisions.
     Steps: you: type /grill-me -> /grill-me (opus · high) -> /grilling (opus · high) -> you: answer each round -> /consult-plan (opus · high) -> you: fits, conflicts or extends
     You decide: every round's answer; what happens to the idea.
  2. grill-docs (skill, from your collection): the same interview, recorded to skilleddocs/ as it settles.
     Steps: /grill-docs (opus · high)
  Run option 1:
  /model opus
  /effort high
  /grill-me a plugin system for the CLI
> /find-me-a-skill
⏺ No task in the prompt or the conversation: reading the board.
⏺ Bash(gh project list --owner @me --format json)
  ⎿ myrepo Board (#3)
⏺ Bash(gh project item-list 3 --owner @me --format json)
  ⎿ Todo: #31 Security pass on the auth refactor branch (size M), #32 Rewrite the sync engine (size L)
⏺ #31 Security pass on the auth refactor branch
  1. security-review (skill, built-in): reviews the branch's pending changes for security issues.
     Steps: /security-review (opus · high, estimate)
  2. gap-scan (skill, from your collection): ranked, evidence-cited gaps report to turn into issues.
     Steps: /gap-scan (opus · high) -> you: pick the items worth doing
  #32 Rewrite the sync engine
  1. divide (skill, from your collection): size L, too big for one session; splits it into session-sized issues.
     Steps: /divide (opus · high, up from medium for size L) -> you: approve the split
  Run for #31:
  /model opus
  /effort high
  /security-review

Related

• next: the board lookup it reuses when you give no task.
• pair and meta-orchestrator: what most implementation suggestions end up pasting.
• grilling: the usual pick when the task isn't clear enough to choose a skill for.

Prereqs

None for a task in words. Board mode needs gh with the project scope, as next does.