| name | animus-skill-authoring |
| description | Build Animus skills - YAML skill definitions, prompts, tool policies, capabilities, adapters, and registries. Use when creating or updating Animus skill files under user or project skill definitions. |
| user_invocable | true |
| auto_invoke | true |
| animus_version | 0.7.0-rc.18 |
Skill Authoring
Skills control how Animus agents behave: prompts, tool policies, model preferences, MCP access, and capability flags.
Do not memorize the full schema before making progress. Start with a narrow edit, then open only the reference that matches the gap:
Minimal pattern
Start from a small skill and add only the sections the task needs:
skills:
code-review:
name: code-review
description: Review code for correctness and risk.
prompt:
directives:
- Flag correctness issues first.
tool_policy:
allow: ["Read", "Grep", "Glob"]
deny: ["Write", "Edit", "Bash"]
capabilities:
is_review: true
writes_files: false
Authoring flow
- Decide whether the skill is primarily about prompt shaping, tool access, model selection, or capability flags.
- Scaffold with
animus skill create --name <slug> --description "..." --prompt "..." (project scope by default; --user for cross-project), or start with the smallest viable YAML shape.
- Add
prompt.directives before reaching for a full prompt.system override.
- Keep tool policies tight and explicit.
- Add adapters only when the behavior must differ by CLI tool.
- After a workflow run, verify the skill actually applied with
animus output phase-outputs --workflow-id <id> — its Skills block shows requested vs applied (with contribution kinds) vs missing.
Where skills apply
- Workflow phases —
phases.<id>.skills and agents.<id>.skills resolve daemon-side at dispatch (union, phase entries first) and ride to the workflow runner as ANIMUS_PHASE_SKILLS_JSON (runner >= v0.4.2; animus daemon preflight warns on older runners). A missing skill name is a loud warning plus a missing metadata record — never a hard failure. Typo'd explicit skills: names also warn at animus workflow config validate / compile time.
- Ad-hoc runs —
animus agent run --skill <name> and animus chat send --skill <name> apply the FULL skill: prompt prefix/suffix/directives wrap the prompt, prompt.system rides the session system prompt, extra_args / env / codex_config_overrides graft onto the launch, and the skill's model / timeout_secs apply when not explicitly given. An unknown --skill name is an error here. Precedence: explicit CLI flags / context-json > skill > defaults; a caller-supplied --runtime-contract-json disables skill application entirely.
Source precedence on both paths: project > user > installed > agent-host (prompt-text-only).
Rules
- Keep one skill focused on one behavior.
- Prefer composable
directives over large system prompts.
- Restrict destructive tools by default unless the skill is explicitly mutating.
- Put project-specific skills in
.animus/config/skill_definitions/.
- Put personal preferences in
~/.animus/config/skill_definitions/.
- Tag skills for discoverability only when tags add real value.
If the task expands into registry operations or a complex schema question, open the matching reference file instead of reading all references preemptively.