Create a Paperclip Bundled Skill

Take source material — a tweet, a task description, a blog post, "make a skill that does X" — and land it as a skill in the Paperclip skills catalog (packages/skills-catalog/), delivered as a reviewed PR. The catalog is the shelf every Paperclip company browses and installs from, so the bar is: correct metadata, useful instructions, worked examples, and a clean validation run.

The core rule is FIND before MAKE: if a good skill already exists (in the catalog, in this repo, or published on GitHub), reference or adapt it instead of writing a duplicate from scratch.

When to use

• A human sends a tweet/link/idea and asks for it to become a Paperclip skill.
• A task asks to bundle an existing repo skill into the catalog.
• A task asks to add an external published skill to the catalog.

When not to use

• The skill is company-private (belongs in that company's library via the Skills UI/API, not the shipped catalog).
• You only need a repo-internal agent skill for working on Paperclip itself — that goes in .agents/skills/ or skills/, with no catalog machinery.

Step 0 — Capture the source material

Understand exactly what the skill should teach before writing anything.

Tweets / X links. Use the xc CLI (X API client). Paperclip engineering agent environments ship it preinstalled and pre-authenticated; it is not a tool you install or mint credentials for yourself. Check availability before relying on it:

command -v xc && xc whoami    # on PATH and authenticated? if not, use the fallback below

xc get <post-url-or-id> --json        # the post itself (conversation_id, author)
xc search 'conversation_id:<id>' --archive --json   # rest of the thread (>7 days old needs --archive)
xc user <username>                    # author context
xc search '<topic keywords>' -n 30    # related discussion

If xc is not on PATH, is unauthenticated, or the account lacks read access (the check above fails for any reason), delegate the fetch to a teammate with X/Twitter access (e.g. the Content Strategist agent) via a child issue: give them the URL and ask for full text of the post + thread

• any linked content.

Other sources. Fetch linked articles/READMEs directly. Record the source URL — it goes in the skill body or PR description as attribution.

Distill: what is the repeatable procedure? What inputs does it take? What does "done" look like? If the source is just an aspiration ("agents should write better commit messages"), you are authoring the procedure yourself — say so in the PR.

Step 1 — FIND: search for an existing skill

Search in this order; stop when you have a clear winner.

1. Already in the catalog? Avoid duplicates (duplicate slugs fail the build):

   grep -i '<topic>' packages/skills-catalog/generated/catalog.json
   ls packages/skills-catalog/catalog/{bundled,optional}/*/
   

2. Already in this repo? Check .agents/skills/, skills/, and issue history (gh search issues / Paperclip board) for prior work on the topic.
3. Published on GitHub? Skills are conventionally a directory with a SKILL.md:

   gh search code --filename SKILL.md "<topic>" --limit 20
   gh search repos "<topic> skill" --limit 20
   

   Also check known collections (e.g. anthropics/skills) and do a web search for <topic> agent skill SKILL.md.

Judge candidates by: does the SKILL.md actually contain the procedure (not a stub)? Is it maintained? What does it bundle (scripts raise the trust level)? Is the license compatible with redistribution? Then pick a path:

• Good external skill exists → add it as an external reference (Step 2A). It stays attributed to and pinned at the upstream repo.
• Partial match → author a local skill (Step 2B) that adapts the idea; credit the source with a link in the SKILL.md body.
• Nothing usable → author a new local skill (Step 2B).

Step 2 — Choose kind, category, and slug

• kind: default to optional. Use bundled only when the skill should ship to every Paperclip company by default — that needs explicit human/board direction, not your judgment call.
• category: reuse an existing directory when one fits (browser, content, docs, finance, paperclip-operations, product, quality, research, software-development). New categories are allowed but must be lowercase kebab-case slugs.
• slug: lowercase kebab-case (^[a-z0-9]+(-[a-z0-9]+)*$), unique across the whole catalog (both kinds).

The skill lives at packages/skills-catalog/catalog/<kind>/<category>/<slug>/ and its canonical key is paperclipai/<kind>/<category>/<slug>.

Step 2A — External reference path (catalog-ref.json)

The directory contains only catalog-ref.json (a directory with both catalog-ref.json and SKILL.md fails the build). The manifest builder fetches the pinned files from GitHub at build time and inventories them.

# Pin the exact commit for the chosen ref (tag or branch)
gh api repos/<owner>/<repo>/commits/<ref> --jq .sha

{
  "source": {
    "type": "github",
    "hostname": "github.com",
    "owner": "<owner>",
    "repo": "<repo>",
    "ref": "<tag-or-branch>",
    "commit": "<40-char sha from above>",
    "path": "<dir inside the repo containing SKILL.md, or ''>"
  },
  "files": ["SKILL.md", "references/**", "scripts/run.py"],
  "defaultInstall": false,
  "recommendedForRoles": ["researcher"],
  "requires": ["python3"],
  "tags": ["topic", "keywords"]
}

Rules the builder enforces:

• files entries are exact relative paths or dir/** globs; SKILL.md must be included and must have frontmatter with name and description.
• If the upstream frontmatter declares key/slug, they must match the catalog placement — otherwise pick a matching slug or use the local path.
• commit must be a full 40-hex SHA; every listed file must be ≤ 1 MiB.
• recommendedForRoles, requires, tags live in the JSON (there is no local SKILL.md to carry them).

See catalog/optional/research/last30days/catalog-ref.json for the live example, and examples/external-reference.md next to this skill.

Step 2B — Author a local catalog skill

Layout:

catalog/<kind>/<category>/<slug>/
├── SKILL.md          # required entrypoint
├── examples/         # 1–2 worked examples (Step 3)
├── references/       # optional deep-dive docs
├── scripts/          # optional — raises trust level, avoid unless needed
└── assets/           # optional templates/images

SKILL.md frontmatter (all validated by the builder):

---
name: <slug>
description: >
  40–300 chars. Routing logic, not marketing: what it does, when to use it,
  when not to.
key: paperclipai/<kind>/<category>/<slug>
recommendedForRoles:
  - engineer            # non-empty; used for staffing suggestions
tags:
  - topic               # non-empty; used for browse/search
---

Optional frontmatter: defaultInstall: true (only for skills every new company should get), requires: [node, python3, ...] for runtime deps.

Body: follow docs/guides/agent-developer/writing-a-skill.md — "When to use" / "When not to use" sections, concrete commands over prose, supporting detail in references/. If the skill came from a tweet or external source, link it in the body for attribution.

Trust level is derived from files, not declared: any scripts/ file makes the skill scripts_executables (install becomes audit-gated and you must extend the scriptBearing expectation in src/shipped-catalog.test.ts); assets/ or non-markdown files make it assets; markdown-only skills stay markdown_only. Prefer markdown-only.

Step 3 — Write 1–2 worked examples

Create examples/ inside the skill directory with one or two markdown files, each a complete input → application → output walkthrough (realistic input, the skill's steps applied, the finished artifact). These ship with the skill so installers can judge it before running it, and they keep the trust level at markdown_only because they are .md files.

Name them by scenario, e.g. examples/rewrite-release-note.md.

Step 4 — Regenerate the manifest and update tests

Never hand-edit generated/catalog.json; it is deterministic build output.

pnpm --filter @paperclipai/skills-catalog build:manifest   # regenerates generated/catalog.json
pnpm --filter @paperclipai/skills-catalog validate         # must report no errors

(External references need network access to GitHub during these steps.)

Then update packages/skills-catalog/src/shipped-catalog.test.ts:

• add the new key to EXPECTED_BUNDLED_KEYS or EXPECTED_OPTIONAL_KEYS (alphabetical order);
• if the skill bears scripts, add it to the scriptBearing expectation.

pnpm --filter @paperclipai/skills-catalog test

The test suite also enforces the ≤300-char frontmatter description budget across the repo and the ≥40-char description / non-empty roles+tags rules for every catalog skill.

Step 5 — Open the PR

Follow the prepare-paperclip-pr skill (.agents/skills/prepare-paperclip-pr/) against paperclipai/paperclip master. The diff should contain exactly:

1. the new skill directory (SKILL.md + examples/ + supporting files, or catalog-ref.json),
2. the regenerated generated/catalog.json,
3. the shipped-catalog.test.ts expectation update.

In the PR body: link the source material (tweet URL, upstream repo), state whether this is a new skill / adaptation / external reference, and note the trust level. Reference PR #10410 (simplified-english) as the shape of a minimal optional-skill PR.

Gotchas

• generated/catalog.json staleness is a validation error — always rerun build:manifest after any file change inside the skill directory (the inventory carries per-file sha256 hashes).
• Duplicate slug across bundled and optional fails the build, not just duplicate keys.
• Symlinks inside a skill directory must resolve within it; directory symlinks are rejected — copy files in.
• The bundled kind and defaultInstall are independent axes; don't set defaultInstall: true casually on optional skills.
• For external references the builder fetches from GitHub on every manifest build; a moved/deleted upstream breaks the build, which is why commit is pinned — prefer upstream tags for ref.