| name | agent-capability-map |
| version | 1 |
| description | Use this skill when the Agent needs to decide where a capability, rule, workflow, runtime anchor, script, job, document, persona change, plugin asset, or repository maintenance item should live. It provides a layered map of the Agent's durable assets so requests to “落实到 all”, “整理自身框架”, or “细化自身” do not bloat memory or scatter files. |
| allowed-tools | read_file list_directory |
Agent Capability Map
Purpose
Provide a compact body map for Agent assets so durable knowledge lands in the right layer instead of bloating memory or scattering across directories.
Use When
- “落实到 all”
- “整理自身框架” / “细化自身”
- “这个应该放哪里”
- “记忆、技能、脚本、job 怎么分”
- “不要一股脑塞进 memory”
- “参考人类构造补能力模块”
Layer Map
memory/: high-frequency global rules loaded every session: identity, safety, durable user preferences, global routing, verification boundaries. Never store secrets, one-time logs, transient state, or long checklists.skills/: reusable workflows, checklists, troubleshooting sequences, routing disciplines, and self-improvement procedures. Use when the content answers “how should I do this kind of task?”scripts/: deterministic repeatable commands or maintenance actions. Scripts should be minimal, documented by an owning skill/job, and safe around secrets.jobs/: scheduled or deferred work. Each job owns one directory with JOB.md and follows the pending/in_progress/completed/cancelled lifecycle.runtime/: non-secret current anchors and local state such as paths, repository mapping, active runtime notes, or generated state that should not be loaded as memory.docs/archive/: long-form references, old plans, investigation notes, or historical documents not needed every turn.activity/: recent interaction history; read-only for the Agent and never manually written.personas/: speaking style only. Persona never overrides MoviePilot identity, safety, workflow, or verification./config local plugin libraries: user-owned MoviePilot plugin source, configs, release assets, and development material when explicitly requested.- MoviePilot tools / slash commands: operational reflexes. Prefer existing commands, MCP tools, plugin commands, schedulers, and workflows before inventing new mechanisms.
Placement Decision Tree
- Is it an external skill/corpus lesson? -> first run
skill-architecture-governance harvest gate; do not import raw directories.
- Must it be loaded every session? ->
memory/
- Is it a reusable procedure, checklist, or routing discipline? ->
skills/
- Is it deterministic executable logic? ->
scripts/
- Is it scheduled or deferred? ->
jobs/
- Is it current non-secret local state, workstream status, candidate inventory, or an anchor? ->
runtime/
- Is it historical or long-form reference? ->
docs/archive/
- Is it only speaking style? -> persona definition
- Is it local plugin development material under
/config? -> local plugin library
- Is it one-time conversation history? -> no manual asset; activity handles it
For large upstream skill packs, prefer: source inventory in runtime/, distilled governance in existing skills/, and long references in docs/archive/. memory/ is the rarest destination.
Verification
After changing a capability asset:
- re-read the changed file
- check frontmatter for skills/jobs
- verify directory-name conventions
- run a minimal assertion when possible
- recommend repository sync for
/config/agent capability changes
Output Contract
State the selected layer, reason, assets changed or intentionally untouched, and verification evidence if changes were made.
