boost

Appearance

Skill assets ​

A skill is not always one file. A skill that runs a script, quotes a long reference, or ships a template needs those files to travel with it, into every agent directory, in every project that receives the skill.

Put them in the skill's own directory. boost sync copies them verbatim beside each emitted SKILL.md.

.ai/skills/codex-review/
├── SKILL.md
└── scripts/
    └── run-codex-review.mjs

After a sync, every selected agent has the whole directory:

.claude/skills/codex-review/
├── SKILL.md
└── scripts/
    └── run-codex-review.mjs

The skill body can then point at its own file, and the path resolves the same in every agent's copy.

What counts as an asset ​

An asset is every file under the skill directory that is not the entry file. Subdirectories keep their structure, so scripts/run.mjs stays at scripts/run.mjs.

Four rules decide the edge cases:

  • The entry file is not an asset. A top-level SKILL.* file is the skill itself. That also protects a file parked beside it: SKILL.md.license is treated as an entry candidate, not shipped as an asset.
  • A deeper SKILL.* file is an asset. examples/SKILL.md is example content, not a second skill.
  • Hidden files are skipped, along with backup and editor temp files, matching how the loader filters skill sources.
  • A flat skill has no assets. skills/my-skill.md has no directory of its own, so there is nothing to collect. Use the nested layout when a skill needs to ship files.

Assets from every source ​

Assets come through all three skill sources: your own .ai/skills/, a vendor package's resources/boost/skills/, and a remote GitHub source. The same collector runs on each, so a skill you install from a package arrives with its scripts intact.

Assets are not executable ​

boost never sets an executable bit. The writer creates every file with the default mode and calls chmod nowhere, so a script arrives at 644 even when the source copy is executable.

Write the skill body to call the interpreter rather than the file:

bash
node scripts/run-codex-review.mjs --uncommitted   # works
./scripts/run-codex-review.mjs --uncommitted      # permission denied

Keep them small ​

Every asset is copied once per selected agent. A 2 MB reference file with six agents selected is 12 MB in the repository, regenerated on every sync. Link out to large material instead of shipping it, and keep in the skill directory what the agent reads or runs.

Ownership ​

The manifest records each asset under its own per-agent path, the same as any other generated file:

json
".claude/skills/codex-review/scripts/run-codex-review.mjs": { ... }

So an asset is reaped when its skill stops shipping, and a file another tool wrote into the same directory is left alone. See File ownership.