Skip to content
AgentsMesh

Installing Anthropic-style skill packs

Anthropic-style skill packs are repos shaped around top-level skills/<kebab>/SKILL.md, optional agents/, references/, and per-target command directories (.claude/commands/, .gemini/commands/). AgentsMesh classifies these automatically and imports them as a single bulk set — no --as or --target needed.

This guide walks the full install → list → uninstall round-trip using addyosmani/agent-skills as the example.

1. Install

Terminal window
agentsmesh init                 # if not already set up
agentsmesh install https://github.com/addyosmani/agent-skills

AgentsMesh:

  1. Acquires .agentsmesh/.install.lock.
  2. Clones the source.
  3. Runs the multi-signal classifier. For agent-skills the classifier matches skills/<kebab>/SKILL.md, agents/<name>.md, references/, multi-tool rules at root, and per-target command dirs — total score ≥ 1.4, primary signal present → anthropic-skill-pack.
  4. Aggregates 23 skills + 3 agents + 7 commands (4 .claude/commands/ + 3 .gemini/commands/ merged into the canonical commands/ set, with claude precedence on basename collisions) + 1 root rule.
  5. Surfaces the bulk-select prompt ([a/n/s]) unless --force.
  6. Materialises the pack under .agentsmesh/packs/addyosmani-agent-skills/ with a per-file integrity manifest at .agentsmesh-install-manifest.json.
  7. Writes the entry to installs.yaml.
  8. Runs agentsmesh generate so every active target (claude-code, cursor, opencode, …) gets the skill files in its native layout.
✓ Installed pack "addyosmani-agent-skills" to .agentsmesh/packs/.
Generated: 67 created, 0 updated, 0 unchanged

2. Inspect

Terminal window
agentsmesh installs list
NAME                     SOURCE                                  FEATURES                  INSTALLED
addyosmani-agent-skills  github:addyosmani/agent-skills@5b4c6da  skills, agents, commands  2026-05-16
Terminal window
ls .claude/skills | head        # writes for the claude-code target
ls .cursor/skills | head        # writes for the cursor target
ls .opencode/skills | head      # writes for the opencode target

Each target’s tree carries exactly the 23 skills, with no boilerplate (README.md, etc.) bleeding through — the install-layer boilerplate filter excludes them.

3. Handle prompts (interactive only)

Bulk-select

When the classifier picks anthropic-skill-pack and stdin is a TTY, the prompt shape depends on how many entities the source carries.

One entity total — single yes/no:

Found in source-name:
  - 1 skill


Install skill "interview-me"? [y/N]

y installs it, n or Enter skips, anything else aborts (exit 130). The three-tier picker would be pointless ceremony here, so AgentsMesh skips it.

Two or more entities — three-tier picker:

Found in agent-skills:
  - 23 skills
  - 3 agents
  - 7 commands
  - 1 root rule


Install [a]ll, [n]one, or [s]elect per type? [a/n/s]
  • a accepts everything (also the default under --force, --json, non-TTY).
  • n returns nothing and exits 0 (no writes).
  • s enters Tier 2, asking once per non-empty kind: Install all 23 skills? [y/n/c] — y/n take/skip the whole kind; c drills into Tier 3 per-entity prompts: Install skill "interview-me"? [y/N/a/q] (a = accept this and every remaining entity in the kind; q = skip this and every remaining entity in the kind).

See the install reference for the full tier matrix, marketplace and multi-collection flows, conflict handling, and the consolidated bypass table.

If a skill body references a file outside its scope (../references/orchestration-patterns.md from a skills/<x>/SKILL.md), you get a per-entity cluster:

Entity "code-reviewer" (agent) has 2 broken links:
  - ../references/orchestration-patterns.md (resolvable-outside)
  - ../references/missing.md (unresolvable)


Action: [i]nclude resolvable as supporting files / [l]eave with warnings / [a]bort install
  • i copies the resolvable file into the entity’s supportingFiles and rewrites the body link to ./references/<x>.md. Unresolvables are left in place with warnings. Bodies authored with the {baseDir}/x.md placeholder or Windows-style ..\refs\x.md rewrite verbatim — the scanner matches the raw destination, not a normalized form.
  • l (the --force default) preserves all link strings and warns.
  • a aborts the run with exit 130; no writes.

When two resolvable links share a basename (docs/A/README.md and docs/B/README.md both pulled in), AgentsMesh slugs the second to keep distinct files: you end up with references/docs-A-README.md and references/docs-B-README.md rather than one file overwriting the other.

4. Uninstall

Terminal window
agentsmesh uninstall addyosmani-agent-skills

AgentsMesh:

  1. Acquires .agentsmesh/.install.lock.
  2. Reads installs.yaml to find the entry.
  3. Reads .agentsmesh-install-manifest.json and computes drift against the current pack contents. When drift is present (locally edited files), the modification prompt offers [d]elete-anyway / [k]eep-modified / [a]bort.
  4. rm -rfs .agentsmesh/packs/addyosmani-agent-skills/.
  5. Drops the installs.yaml row (and any matching extends: row).
  6. Runs agentsmesh generate so cleanupStaleGeneratedOutputs evicts orphaned target files (.claude/skills/<x>/, etc.).
Terminal window
agentsmesh installs list
# No installed packs.

--keep-pack preserves the directory on disk but still drops the yaml rows (useful when you’ve edited files and want to keep them around). --keep-generated skips the post-uninstall generate; target files stay until the next manual generate — a warning lists what will be cleaned then.

Partial-batch failures

agentsmesh uninstall name1,name2,name3 --force (or --all) wraps each pack in its own try/catch. If one pack’s apply step throws (EACCES, disk full, etc.), survivors still apply, post-operation generate still runs over them, and the failed pack lands in the JSON envelope’s data.failed[] array. The run exits 1 whenever data.failed[] is non-empty — see agentsmesh uninstall reference for the JSON shape.

Where to dig deeper