Find Me Skills

Help a user who has a goal but not a skill list find the right Agent Skills, explain what each one does, lay out an order to run them in, and — if they approve — hand them a single installable bundle file.

The user's defining trait is that they don't know what to ask for. Someone who already knows they want frontend-design should just run asm install …. This skill is for "I'm building an app and want to do marketing from scratch, but I don't know marketing or which skills exist." Draw out the goal, confirm it, then map it onto real, installable skills from the live catalog.

The loop

1. Collect intent — conversationally draw out what the user is trying to achieve.
2. Confirm understanding — play back your read of their situation; let them correct it before you search.
3. Discover — query the live asm catalog for candidate skills (never guess skill names).
4. Curate — dedupe, group by step, and explain each skill in one plain sentence.
5. Sequence — give a step-by-step path with the input and output of each step.
6. Export — on approval, write a bundle file and give the one-line install command.

Start at the step the user is already at. If they open with a rich goal ("I need SEO, a landing page, and launch copy for my SaaS"), confirm quickly and move to discovery. If they're vague ("help me market my app"), spend more time in steps 1–2. Never skip step 2: confirming before searching keeps recommendations relevant.

If the user asks for one capability ("is there a skill that converts PDFs?"), skip the loop: run one asm search "<term>" --available --json and give the matching asm install <installUrl> line.

Prerequisite (check once, up front)

This skill drives the asm CLI for discovery and produces a file it installs. Run this before promising recommendations:

command -v asm || echo "MISSING"

If the output is MISSING, tell the user the skill needs the Agent Skill Manager CLI on PATH, point them at npm install -g agent-skill-manager, and stop.

Step 1 — Collect intent

Ask open questions, one or two at a time, until you can state the user's goal in a sentence. Useful prompts:

• What are you building or working on right now?
• What outcome do you want — a launched product, a written artifact, a faster workflow?
• What part feels hardest or most unfamiliar? (This is often where skills help most.)
• Is this a one-off task or something you'll repeat?

Match your vocabulary to theirs. A non-marketer asking for "marketing" may actually need positioning, a landing page, and launch copy — offer those as options; do not assume. Avoid jargon ("ICP", "ASO") unless they use it first.

Step 2 — Confirm understanding (do not skip)

Before searching, play back what you heard and ask for an explicit confirmation:

Here's what I understand: you're building {project}, and you want to {goal}. The pieces you're unsure about are {gaps}. Did I get that right?

If they correct you, fold it in and ask again. Move on only after they agree. Confirming the situation is what prevents a confidently-wrong skill list.

Step 3 — Discover candidates from the live catalog

Never invent skill names or install URLs. The catalog changes constantly; the only trustworthy source at runtime is the asm CLI on the user's machine. Derive 2–5 search terms from the confirmed goal and query each:

asm search "<term>" --available --json

Run a separate search per term — broad terms ("marketing", "seo", "landing page", "launch") surface different skills. Read references/catalog-discovery.md when you need the JSON shape, installed-vs-available rules, or empty-result handling; keep this detail out of the main context budget until Step 3 needs it.

Read each candidate's description to judge relevance; its "Use when…" / "Don't use for…" text tells you whether it fits the goal. Then run asm search "<term>" --json without --available to detect installed skills. Mention installed matches in the plan, but exclude them from the bundle. Only available skills with an installCommand go in.

Step 4 — Curate: dedupe and explain

From the union of search hits, build the recommendation set:

• Deduplicate by skill name. Keep one entry per name. If two repos offer the same name, keep the one whose description best matches the goal and note the choice in the plan.
• Drop weak fits. Keep only skills you can justify in one sentence tied to the user's goal.
• Explain each in plain language — one sentence on what it does for this user, not a paraphrase of its description. "landing-page-copywriter writes the words for your launch page so visitors understand and sign up."

Step 5 — Sequence into a step-by-step path

Order the curated skills into the sequence the user should run them in. For each step, state its input (what the user or the previous step provides) and its output (what they'll have after). Foundational/context skills usually come first; review/QA skills usually come last. Example shape:

Step 1 — marketing-context
  in:  your product description, target customer
  out: a saved brand/positioning brief other skills read first
Step 2 — landing-page-copywriter
  in:  the brief from step 1
  out: landing-page copy ready to paste

Show this plan to the user before exporting anything. Make it easy to say "drop step 2" or "add something for email". After any change, show the revised plan and ask for approval again.

Step 6 — Export an installable bundle (on approval)

Only after the user approves the plan, write a bundle file in asm's BundleManifest format with a goal-based name such as marketing-starter.bundle.json. Write it to the current directory unless the user names another path.

See references/bundle-format.md for the required JSON template, validation rules, and the reason this skill uses asm bundle install instead of asm install or asm import. Copy each installUrl from Step 3 verbatim; never hand-construct it and never include already-installed skills.

Check the file before handing it off:

asm bundle show ./marketing-starter.bundle.json --json

• If the command exits 0 and its skills[].name list matches the approved plan, give the user the install command asm bundle install ./marketing-starter.bundle.json. In a terminal it asks for the tool, then which bundle skills to install, then the scope. Outside a terminal, -p/--tool <tool> is required; -y skips the skill and scope pickers.
• If the command errors, fix the field the error names, then run the check again. After two failed checks, stop, report blocked with the error, and do not hand off an install command.

Step Completion Reports

After each step, print a compact status block: √ pass, × fail, — context.

◆ Step N — [step name]
··································································
  [check]:           √ pass
  Result:            PASS | FAIL | PARTIAL

Checks per step:

• Prerequisite — asm on PATH
• Steps 1–2 — goal stated in one sentence, user confirmed the goal
• Step 3 — one available search per term, one installed search per term
• Steps 4–5 — no duplicate names, every step has in/out, user approved the plan
• Step 6 — bundle file written, asm bundle show --json exit 0, listed skills match the plan

Final output

End every run with a short report the user can read without scrolling back:

1. Result — the first line states complete, plan only (nothing to bundle, or the user declined export), or blocked (with the reason).
2. The plan — numbered steps, each with the skill, a one-line purpose, and in/out.
3. Evidence — the search terms you ran, the bundle file path, and whether asm bundle show passed.
4. Uncertainty — say that each recommendation is based on the skill's catalog description; you did not run, test, or security-audit any skill. Name any part of the goal with no catalog match.
5. Decision — for complete, the user's next action: asm bundle install ./<file> on its own line; installation happens only when they run it, so no other approval is needed. For plan only or blocked, name what would unblock the next step, or state that no action is needed.

Keep explanations plain. The user came here because they didn't know the landscape — leave them understanding what they're about to install and why.

Acceptance Criteria

Verify all of these before calling the run complete:

• The user explicitly confirmed the goal before any catalog search.
• At least one asm search "<term>" --available --json query was run for each chosen search term.
• Installed skills were checked with asm search "<term>" --json and excluded from the bundle.
• The user approved the sequenced plan before any file was written.
• Expected output for complete: a numbered plan, a bundle file path, and an install command. For plan only: the numbered plan and no bundle file.
• For complete, asm bundle show ./<file>.bundle.json --json succeeded; otherwise the run reports blocked with the validation error instead of handing off a broken command.

When reviewing a run's output (by eval or by a human), also check that a reader can:

• Find the main result — the first line states the completion status.
• Separate facts from assumptions — catalog-description judgements and untested behavior are labeled as such.
• Trace claims — each recommended skill maps to a search term and a catalog result; "bundle valid" maps to the asm bundle show check.
• See the next decision — the install command is named, or the report states why there is none.

An agent's own review cannot confirm human understanding. A reviewer who received no human feedback records understanding as unconfirmed.

Edge cases

• asm not installed — stop at the prerequisite check; point them at the install docs.
• Vague goal that won't sharpen — stay in steps 1–2; offer 2–3 concrete directions ("Do you mean A, B, or C?") rather than searching on a guess.
• No catalog matches for part of the goal — say so; recommend only what fits, and suggest skill-creator if they may need to author something that doesn't exist yet.
• Everything relevant is already installed — there's nothing to bundle; give the step-by-step plan using their installed skills, skip the export, and report the result as plan only.
• User declines the plan — don't write a file. Adjust from their feedback and ask again, or stop and report plan only.
• Duplicate skill names across repos — keep one; pick the better-matching description and note the choice.