Manage Agents

Manage Agents & Tooling

Model policy + roster: .claude/system/registry.md. Update its table on EVERY create/remove, and progress_log the change.

Create an agent

1. Check .claude/agents/ — don't duplicate an existing specialty; sharpen the existing one instead.
2. Copy .claude/system/templates/agent-template.md → .claude/agents/.md. Fill: name, description ("Use for …" so auto-selection works), model per registry policy, 5–8 specialty bullets that would actually change how the work is done.
3. New agent TYPES register at the next session start. Until then, spawn general-purpose with the file body prepended + the model param — same file, either path.

Modify an agent

Edit .claude/agents/.md directly — description, model tier (justify against registry policy), specialty bullets. Typical at init: retuning a generic role to the project's domain (e.g. backend-engineer → "REST + Postgres" bullets swapped for "gRPC + embedded sqlite"). Sync the registry row. Named-type changes load next session; immediate need → general-purpose + updated file body.

Remove an agent

Delete .claude/agents/.md + its registry row. Never remove a role with in-flight tasks — reassign first. If the user created it manually, ask before deleting.

Create a skill

.claude/skills//SKILL.md — frontmatter (name, description with a "Use when …" trigger), body = the procedure. Worth creating when a procedure has ≥3 steps and will run ≥2 times. Loads next session; follow the file directly until then.

Hooks

Scripts live in .claude/hooks/; register in .claude/settings.json under hooks. (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop):

{"hooks": [{"type": "command", "command": "bash \"$CLAUDE_PROJECT_DIR/.claude/hooks/.sh\""}]}

Keep hooks fast (": {"command": "...", "args": ["..."]}}}. Prefer stdio + zero/low-dependency (see memory-server.mjs as the pattern). Loads next session. **NEVER remove the memory` server** — the system is amnesiac without it.

Removal safety (all artifact types)

State why, progress_log it, and if the artifact was user-added rather than system-created, ask the user first.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zdnemz
• Source: zdnemz/claude-team
• License: MIT

Install and usage instructions live in the source repository linked above.