Swarm skill forge
A shared knowledge vault + full software-engineering workflow for AI coding agents — run Claude Code and Codex in parallel with one memory, one plan.
npx -y skills add AnmarHani/SwarmVault --skill swarm-skill-forgeAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 20 days oldThe repository was created 20 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Author new high-quality skills that trigger reliably and stay lean. Use when creating or editing a skill, packaging a workflow as a skill, or deciding whether something should be a skill at all.
SKILL.md
2.9 KB, 663 tokens by cl100k_base, as published. Nobody here has run it
swarm-skill-forge — skill authoring
A skill is packaged judgment: instructions a future agent follows instead of its defaults. Write one only when that trade is worth it.
Gate: should this be a skill?
- Recurring task with a non-obvious right way → skill.
- Single fact or preference → memory note. Project convention → CLAUDE.md/AGENTS.md line.
- One-off → nothing. When in doubt, don't — catalogs rot.
- Overlap check first:
lsthe installed skills; if an existing skill covers 70% of the job, extend it instead of adding a sibling (two skills for one job = router confusion).
Structure
---
name: kebab-case-verb-or-domain
description: <what it does> + <every trigger phrasing a user might use>. The
description IS the router — write it for the model deciding whether to load.
---
# Name — one-line job
Gate / when NOT to use.
Numbered steps, imperative, each observable ("run X, expect Y").
Rules the defaults would get wrong.
Worked example (input → output).
- Platform-agnostic markdown only — it must read the same via Claude Code skills and a Codex AGENTS.md index.
- Portable on its own (NFR-C4): the folder must work when copied out alone, with no
vault, no
swarmvault.py, and no siblings installed. Name peer skills by what they add, never as a prerequisite, and give every vault-backed path a local fallback ("voice.mdin the vault, else.writing/voice.md"). Skills about the vault are the exception. Degrade, don't downgrade: write the vault path first and fullest, then the fallback. A skill weakened so a lone copy can satisfy it has failed this rule, not passed it. - Heavy reference material (> ~150 lines) →
references/*.md, loaded on demand; the SKILL.md tells the reader which reference to load when. - Machine-lane style for checklists/tables; prose where nuance matters.
Verify before shipping (skills are code — test them)
- Trigger dry-run: write 3 paraphrases of how a user would ask; does the description clearly fire for them — and clearly NOT fire for near-misses?
- Toy walk: execute the skill on a small real case; every step must be doable without information the skill forgot to mention.
- Altitude check: delete every line the model would do correctly anyway — a skill states only what defaults get wrong.
Ship
Project skills → .claude/skills/<name>/ + an AGENTS.md index line (name, description,
when-to-use, path) for Codex. Contributions to SwarmVault itself → canonical skills/
- a CREDITS.md entry if the technique has an upstream.
Influences: superpowers writing-skills; Pocock's writing-great-skills — see CREDITS.md.