samugit83/redamon/add-community-skill

When to Use

• Shipping a battle-tested attack workflow as an importable .md, without

hardcoding it into Python.

For a hardcoded, first-class skill (9-layer wiring, badge, per-project toggle), use builtin-agent-skill instead. For a reference/theory doc rather than a workflow, that is a Community Chat Skill (different flow).

Critical Rules

• NEVER invent tool names. Every step must name a key of TOOL_REGISTRY

(agentic/prompts/tool_registry.py): query_graph, kali_shell, execute_curl, execute_code, execute_playwright, execute_nuclei, execute_hydra, execute_ffuf, ... A step naming a non-existent tool is dead on arrival.

• NEVER call a `kali_shell` CLI that is not installed AND documented. The

binary must be in mcp/kali-sandbox/Dockerfile and named in the kali_shell description (the kali_toolbox MCP tool serves that text verbatim; an unlisted binary does not exist for MCP clients). Prefer a tool already installed. A new one is an agent-tool change: PROMPT.ADD_AGENTIC_TOOL.md Type A (pinned install, description, stealth rule, smoke test). Every wordlist or file path a command uses must exist in the image too.

• NEVER rebuild the agent for the `.md` alone. ./agentic/community-skills is

volume-mounted read-only into the container (docker-compose.yml:903); the GET /community-skills endpoint auto-discovers by globbing the directory. Drop the file and it is live on the next call. (This is the exception to "agentic/ changes need a rebuild" - that applies to baked Python, not the mounted skills dirs.) A skill that needs a NEW kali CLI is not rebuild-free: rebuild kali-sandbox AND agent.

• NEVER write a description that overlaps a built-in skill. Imported skills

compete with built-ins in the Intent Router; a generic or overlapping opening paragraph means the classifier never selects it. Make the first paragraph and the description distinct. If the skill covers the same attack class as a built-in (KNOWN_ATTACK_PATHS in agentic/state.py), add a vs. <builtin> (built-in) line telling the user to disable that built-in on import.

• Commands must be copy-pasteable in fenced blocks - no pseudo-code. The

markdown content is injected verbatim into the system prompt.

• Remember it is per-user, not global. Users import via Global Settings >

Agent Skills; already-imported users do NOT auto-pick-up new skills (they re-import; duplicates are skipped by name). The same skip means an EDIT to a shipped skill never reaches existing importers: they must delete the old copy and re-import. Say so in the CHANGELOG.

Commands

# No build. Drop the file and confirm discovery:
ls agentic/community-skills/            # your new <skill>.md lives here
# GET /community-skills globs this dir; the row lands in the Postgres UserAttackSkill
# table only after a user clicks "Import from Community".

If agentic/community-skills/README.md has a skills table, add a row (check the file first; do not invent structure).

Resources

• docs/readmes/coding_agent_prompts/PROMPT.ADD_COMMUNITY_AGENT_SKILL.md - full authoring flow, classification tips, troubleshooting
• Related skill: builtin-agent-skill