rules/delegation-brief.md

GitHub

Delegating skill authoring to parallel agents

Delegate when three or more skills need substantial work. One agent per skill is wasteful. Group the work into ≤ 7 agents, with 2–3 skills each, and put overlapping topics in the same agent so that boundaries stay consistent. The first build used these groups:

Agent Skills Why together
A modern-migration, modules-loading both cover module.register EOL and strip-types renames
B histogram-stats, bench, process-timeout bench results relate to Histogram
C test-runner, utilities broad, small APIs
D sqlite, zip, stream-iter zlib iterable compression is shared by zip and stream-iter
E crypto, permissions --allow-openssl-store spans both
F diagnostics, networking channels, profiling, net
G vfs, ffi, quic-dtls experimental modules that need special flags or builds

Do the shared scaffolding yourself before you spawn the agents: tools, template changes, and any new convention. Agents copy whatever exists when they start. A convention added mid-flight, such as the exit-77 rule, has to be sent to every running agent.

Brief template

Fill in the parts in angle brackets. Every line exists because an agent once got it wrong.

You are authoring/updating agent skills in /home/dev/node-skills. They teach coding
agents about Node.js features newer than their training data. FIRST read
AGENTS.md (the authoring contract; follow it exactly) and SURVEY.md. Then read
the finished exemplar skill skills/<a good existing skill>/ to match its style.

## Ground truth
- Docs: <node checkout>/doc/api/*.md at commit <sha> (pin this in meta.json).
  Read the YAML added:/changes: and Stability lines for every API you document.
- Binaries: main build <path> (<version>); releases <oldest v24>, <latest v24>,
  <latest current>.
- Every non-trivial claim must be verified by an example you run. If the docs and
  behavior disagree, document the behavior and record the discrepancy (with
  doc/api file:line) in meta.json "discrepancies" AND in your report.
  If you can't verify something, leave it out or mark it as docs-only.

## Your skills (create or edit only these directories)
1. `<skill>` — <scope: APIs, flags, sources, the user TASKS to focus on, the
   "Don't" list of stale alternatives, boundaries with sibling skills by name>

## Constraints
- Floor is v24 LTS. Version columns use the full YAML added: list.
- The description routes: task first, then triggers, then API names, then the
  version floor; it must be a double-quoted YAML string ≤ 200 tokens of
  frontmatter. Body ≤ 1000 tokens. No "When to use" heading.
- Stability 1.x content needs the verbatim **Experimental API.** banner.
- Examples: line-1 `// @node-skills {...}` header, self-asserting, no network,
  no leftover files. Exit 77 with "skip: <reason>" when the example can't run
  at runtime. Never exit 0 without running the checks.
- Validate: skill-lint, skill-budget --check, node --test tools/, and
  run-examples on EVERY binary listed above.
- Do not edit outside your directories. Run no git commands that modify state.

## Final report (short)
Per skill: files, budget numbers, example results per binary, discrepancies
(file:line), anything left out or unverified.

While the agents run

  • Don't do overlapping work yourself. Use the time for shared files: the README, CI, installers.
  • An agent that finished keeps its context. Send it follow-ups (a new binary, a failing example) through SendMessage instead of spawning a fresh agent.

Reviewing what comes back

  • Re-run the gate yourself. Agents report repo-wide failures caused by other in-progress agents, which is normal mid-flight; the gate must be clean at the end.
  • Read every SKILL.md body: check the version columns, whether a feature is really unavailable on v24, and that nothing is claimed beyond what an example shows.
  • Treat reported discrepancies as leads, and spot-check the surprising ones. In the first build, an agent's report also caught an error in SURVEY.md.
  • Look for exit-0 skips: grep -rn "exit(0)" skills/*/examples.