Manage

Manage the lifecycle of agents and skills in the .claude/ directory. Handles creation with rich domain content, atomic updates (renames) with cross-reference propagation, and clean deletion with broken-reference cleanup. Keeps the MEMORY.md inventory in sync with what actually exists on disk.

• $ARGUMENTS: required, one of:
  • create agent <name> "description" — create a new agent with generated domain content
  • create skill <name> "description" — create a new skill with workflow scaffold
  • update agent <old-name> <new-name> — rename agent file + update all cross-refs
  • update skill <old-name> <new-name> — rename skill directory + update all cross-refs
  • delete agent <name> — delete agent file + clean broken refs
  • delete skill <name> — delete skill directory + clean broken refs
  • add perm <rule> "description" "use case" — add a permission to settings.json allow list and permissions-guide.md
  • remove perm <rule> — remove a permission from settings.json allow list and permissions-guide.md
• Names must be kebab-case (lowercase, hyphens only)
• Descriptions must be quoted when they contain spaces
• Permission rules use the Claude Code format: WebSearch, Bash(cmd:*), WebFetch(domain:example.com)
• --skip-audit — optional flag: skip the Step 9 /audit validation (use when running inside an audit fix loop to avoid recursion)

Agent examples:

• /manage create agent task-planner "Planning specialist for decomposing epics into actionable tasks"
• /manage update agent example-agent example-agent-v2
• /manage delete agent old-agent-name

Skill examples:

• /manage create skill benchmark "Benchmark orchestrator for measuring and comparing performance across commits"
• /manage update skill debug trace-logger
• /manage delete skill old-skill

Permission examples:

• /manage add perm "Bash(jq:*)" "Parse and filter JSON" "Extract fields from REST API responses"
• /manage remove perm "Bash(jq:*)"

• AGENTS_DIR: .claude/agents
• SKILLS_DIR: .claude/skills
• USED_COLORS: blue, green, purple, lime, orange, yellow, cyan, violet, teal, indigo, magenta, pink
• AVAILABLE_COLORS: coral, gold, olive, navy, salmon, red, maroon, aqua, brown

Task tracking: per CLAUDE.md, create tasks (TaskCreate) for each major phase. Mark in_progress/completed throughout. On loop retry or scope change, create a new task.

Step 1: Parse and validate

Extract operation, type, name, and optional arguments from $ARGUMENTS.

Validation rules:

• Name must match ^[a-z][a-z0-9-]*$ (kebab-case)
• For create: name must NOT already exist on disk
• For update/delete: name MUST already exist on disk
• For update: new-name must NOT already exist on disk
• For create: description is required
• For add perm: rule must NOT already exist in settings.json allow list; description and use case are required
• For remove perm: rule MUST already exist in settings.json allow list

Use the Glob tool to check existence:

• Agent: pattern agents/<name>.md, path .claude/ — non-empty result means it exists
• Skill: pattern skills/<name>/SKILL.md, path .claude/ — non-empty result means it exists

# Check permission existence (for add perm / remove perm)
python3 -c "import json,sys; d=json.load(open('.claude/settings.json')); sys.exit(0 if '<rule>' in d['permissions']['allow'] else 1)" 2>/dev/null

If validation fails, report the error and stop.

For perm operations: skip Steps 2, 3, 5, 6, 7, 8, 9 — go directly from Step 1 → Step 4 → Step 10.

Step 2: Overlap review (create only)

Before creating anything, check if existing agents/skills already cover the requested functionality:

1. Read descriptions of all existing agents (head -n 3 of each .md in agents/) and skills (head -n 3 of each SKILL.md)
2. Compare the new description against each existing one — look for domain overlap, similar workflows, or redundant scope
3. Present findings to the user:
   • No overlap: proceed to Step 3
   • Partial overlap: name the overlapping agent/skill, explain what it covers vs what the new one would add, and ask the user whether to proceed, extend the existing one instead, or abort
   • Strong overlap: recommend against creation — suggest using or extending the existing agent/skill instead

Skip this step for update, delete, and perm operations.

Step 3: Inventory current state

Snapshot the current roster for later comparison. Steps 2 and 3 are independent reads — issue Glob calls for both in the same response.

Use Glob (pattern agents/*.md, path .claude/) for agents and Glob (pattern skills/*/, path .claude/) for skills to build the name lists. Use Grep (pattern ^color:, glob agents/*.md, path .claude/, output mode content) to collect colors currently in use.

For name extraction that requires basename/sort string processing, feed the Glob results into Bash:

# Current agents (Glob provides paths; Bash extracts sorted names)
ls .claude/agents/*.md | xargs -n1 basename | sed 's/\.md$//' | sort

# Current skills (Glob provides paths; Bash extracts sorted names)
ls -d .claude/skills/*/ | xargs -n1 basename | sort

Step 4: Execute operation

Branch into one of six modes:

Mode: Create Agent

1. Fetch the latest Claude Code agent frontmatter schema to ensure the template is current:

   • Spawn web-explorer to fetch https://code.claude.com/docs/en/sub-agents with instruction: "Write your full findings (schema fields, new fields, deprecated fields) to /tmp/manage-schema-$(date +%s).md using the Write tool. Return ONLY a summary line: fields=N new=N deprecated=N schema_file=/tmp/manage-schema-<ts>.md"
   • Read the returned summary and use it to extract: valid frontmatter fields (name, description, tools, disallowedTools, model, permissionMode, maxTurns, skills, mcpServers, hooks, memory, background, isolation), current model shorthands, and any new fields
   • Note any new fields worth including in the generated template Adjust the template generated in steps 2–4 to reflect the current schema. If a new field is broadly useful for the agent's role (e.g. maxTurns for long-running agents), include it with a sensible default and inline comment.

2. Pick the first unused color from the AVAILABLE_COLORS pool (compare against colors found in Step 3)

3. Choose model based on role complexity:

   • opusplan — plan-gated roles (solution-architect, oss-maintainer, self-mentor): long-horizon reasoning + plan mode
   • opus — complex implementation roles (sw-engineer, qa-specialist, ai-researcher, perf-optimizer): deep reasoning without plan mode
   • sonnet — focused execution roles (data-steward, web-explorer, doc-scribe): pattern-matching, structured output
   • haiku — high-frequency diagnostics roles (linting-expert, ci-guardian): rule-application, structured lint output

4. Write the agent file with real domain content derived from the description:

Read the agent scaffold template from .claude/skills/manage/templates/agent-scaffold.md and follow the content rules and tool selection guidelines it defines.

Mode: Create Skill

1. Fetch the latest Claude Code skill frontmatter schema to ensure the template is current:

   • Spawn web-explorer to fetch https://code.claude.com/docs/en/skills with instruction: "Write your full findings (schema fields, new fields, deprecated fields) to /tmp/manage-skill-schema-$(date +%s).md using the Write tool. Return ONLY a summary line: fields=N new=N deprecated=N schema_file=/tmp/manage-skill-schema-<ts>.md"
   • Read the returned summary and use it to extract: valid frontmatter fields (name, description, argument-hint, disable-model-invocation, user-invocable, allowed-tools, model, context, agent, hooks), and any new fields
   • Note any new fields worth including in the generated template Adjust the template generated in step 3 to reflect the current schema. Include model or context: fork only when the skill's described purpose clearly benefits from them.

2. Create the skill directory

3. Write the skill file with workflow scaffold:

mkdir -p .claude/skills/<name>

Read the skill scaffold template from .claude/skills/manage/templates/skill-scaffold.md and follow the content rules it defines.

Mode: Update Agent

Atomic update — write new file before deleting old:

1. Read .claude/agents/<old-name>.md using the Read tool.

# 2. Write new file with updated name in frontmatter
# (Edit the `name:` line in frontmatter to use new-name)

# 3. Verify new file exists and is valid
head -5 .claude/agents/<new-name>.md

# 4. Delete old file only after new file is confirmed
rm .claude/agents/<old-name>.md

Mode: Update Skill

Atomic update — create new directory before removing old:

# 1. Create new directory
mkdir -p .claude/skills/<new-name>

# 2. Copy SKILL.md with updated name in frontmatter
# (Read old, edit name: line, write to new location)

# 3. Verify new file exists
head -5 .claude/skills/<new-name>/SKILL.md

# 4. Remove old directory only after new is confirmed
rm -r .claude/skills/<old-name>

Mode: Delete Agent

# Confirm existence before deleting
ls .claude/agents/<name>.md
rm .claude/agents/<name>.md

Mode: Delete Skill

# Confirm existence before deleting
ls .claude/skills/<name>/SKILL.md
rm -r .claude/skills/<name>

Mode: Add Permission

Adds a rule to both settings.json and permissions-guide.md atomically.

1. Determine the guide category from the rule prefix:

   • WebSearch → ## Web
   • WebFetch(domain:...) → ## WebFetch — allowed domains
   • Bash(gh ...) → ## GitHub CLI — read-only
   • Bash(git log:*), Bash(git show:*), Bash(git diff:*), Bash(git rev-*:*), Bash(git ls-*:*), Bash(git -C:*), Bash(git branch:*), Bash(git tag:*), Bash(git status:*), Bash(git describe:*), Bash(git shortlog:*) → ## Git — read-only
   • Bash(git add:*), Bash(git checkout:*), Bash(git stash:*), Bash(git restore:*), Bash(git clean:*), Bash(git apply:*) → ## Git — local write
   • Bash(pytest:*), Bash(python ...), Bash(ruff:*), Bash(mypy:*), Bash(pip ...) → ## Python toolchain
   • Bash(brew ...), Bash(codex:*) → ## macOS / ecosystem
   • All other Bash(...) → ## Shell utilities

2. Update settings.json — parse, append, write back:

python3 -c "
import json
with open('.claude/settings.json') as f:
    d = json.load(f)
d['permissions']['allow'].append('<rule>')
with open('.claude/settings.json', 'w') as f:
    json.dump(d, f, indent=2)
    f.write('\n')
"

3. Update permissions-guide.md — append a new row to the end of the correct section (before its trailing --- separator). New row format:

| `<rule>` | <description> | <use case> |

Use the Edit tool to insert the row: find the last table row in the target section and insert after it.

4. Verify both files were updated:

python3 -c "import json; d=json.load(open('.claude/settings.json')); print('OK' if '<rule>' in d['permissions']['allow'] else 'MISSING')"
grep -F '`<rule>`' .claude/permissions-guide.md

Mode: Remove Permission

Removes a rule from both settings.json and permissions-guide.md atomically.

1. Update settings.json — parse, filter, write back:

python3 -c "
import json
with open('.claude/settings.json') as f:
    d = json.load(f)
d['permissions']['allow'] = [p for p in d['permissions']['allow'] if p != '<rule>']
with open('.claude/settings.json', 'w') as f:
    json.dump(d, f, indent=2)
    f.write('\n')
"

2. Update permissions-guide.md — use the Edit tool to remove the table row containing `<rule>`.

3. Verify both files are clean:

python3 -c "import json; d=json.load(open('.claude/settings.json')); print('OK' if '<rule>' not in d['permissions']['allow'] else 'STILL PRESENT')"
grep -cF '`<rule>`' .claude/permissions-guide.md && echo "STILL IN GUIDE" || echo "OK"

Step 5: Propagate cross-references

Search all .claude/ markdown files for the changed name and update references:

Use the Grep tool to find all references to the name across the config:

• Pattern <name>, glob agents/*.md, path .claude/, output mode content
• Pattern <name>, glob skills/*/SKILL.md, path .claude/, output mode content
• Pattern <name>, file .claude/CLAUDE.md, output mode content
• Pattern <name>, file README.md, output mode content

For update: Use the Edit tool to replace every occurrence of <old-name> with <new-name> in each file that references it.

For delete: Review each reference. If the deleted name appears in:

• A cross-reference suggestion (e.g., "use X agent") — remove or replace with the closest alternative
• An inventory list — remove the entry
• A workflow spawn directive — flag for manual review

For create: No cross-ref propagation needed (new names have no existing references).

Step 6: Update MEMORY.md roster (auto-memory)

MEMORY.md is Claude Code's auto-memory file — it is not stored under .claude/. It is injected into the conversation context at session start. The absolute path appears near the top of the system prompt (e.g. ~/.claude/projects/.../memory/MEMORY.md). Use that absolute path with the Edit tool.

Regenerate the inventory lines from what actually exists on disk:

Use Glob (agents/*.md, path .claude/) for agents and Glob (skills/*/, path .claude/) for skills to get file paths, then feed into Bash for name extraction and comma-joining (no dedicated tool for aggregate string joining):

# Get current agent list (Glob provides paths; Bash joins names)
ls .claude/agents/*.md | xargs -n1 basename | sed 's/\.md$//' | paste -sd', ' -

# Get current skill list (Glob provides paths; Bash joins names)
ls -d .claude/skills/*/ | xargs -n1 basename | paste -sd', ' -

Use the Edit tool with the absolute auto-memory path from the conversation context to update these two roster lines in MEMORY.md — they must stay in sync with disk state:

• - Agents: oss-maintainer, sw-engineer, ... (the roster line, not the path line)
• - Skills: review, survey, ...

Step 7: Update README.md

Update the agent or skill table in README.md:

• create agent: add a new row to the ### Agents table — columns: | **name** | Short tagline | Key capabilities |
• create skill: add a new row to the ### Skills table — columns: | **name** | \/name` | Description |`
• update (rename): find and replace the old name in the table row with the new name
• delete: remove the row for the deleted name

The README tables are self-documenting — keep descriptions concise (one line) and consistent in tone with the surrounding rows. Do not add/remove table columns.

Step 8: Verify integrity

Confirm no broken references remain:

Use the Grep tool to extract all backtick-quoted agent/skill name references (pattern `[a-z]+(-[a-z]+)+`, glob {agents/*.md,skills/*/SKILL.md}, path .claude/, output mode content).

Use Glob (agents/*.md, path .claude/) and Glob (skills/*/, path .claude/) for the on-disk inventory; extract names via Bash if needed for comparison:

# On-disk agent names (Glob provides paths; Bash extracts names)
ls .claude/agents/*.md | xargs -n1 basename | sed 's/\.md$//'
# On-disk skill names (Glob provides paths; Bash extracts names)
ls -d .claude/skills/*/ | xargs -n1 basename

Use Grep to search for the specific changed name and confirm:

• Update: zero hits for old name, appropriate hits for new name
• Delete: zero hits for deleted name (or flagged references noted)
• Create: new file exists with valid structure

For create and update operations, also verify tool efficiency: cross-check the agent/skill's declared tools (tools: or allowed-tools:) against the tool names that actually appear in the workflow body. Any declared tool not referenced anywhere in the content should be flagged as a cleanup candidate (unnecessary permission surface).

Step 9: Audit

Run /audit to validate the created/modified file(s) and catch any issues introduced by this operation. Skip this step if invoked with --skip-audit or if the current manage operation is itself being executed as part of an audit fix run — the outer audit will cover it.

/audit

For a targeted check of only the affected file, spawn self-mentor directly:

• For create: audit the new file for structural completeness, cross-ref validity, and content quality
• For update: audit the renamed file and verify no stale references remain
• For delete: audit remaining files for broken references to the deleted name

Include the audit findings in the final report. Do not proceed to sync if any critical findings remain.

Step 10: Summary report

Output a structured report containing:

• Operation: what was done (create/update/delete + type + name, or add/remove perm + rule)
• Files Changed: table of file paths and actions (created/renamed/deleted/cross-ref updated/appended/removed)
• Cross-References: count of files updated, broken refs cleaned (n/a for perm operations)
• Current Roster: agents (N) and skills (N) with comma-separated names (n/a for perm operations)
• Audit Result: audit findings (pass / issues found) (n/a for perm operations)
• Follow-up: run /sync apply to propagate to ~/.claude/; for create review generated content; for agent create/update/delete run /calibrate routing fast — any change to the agent roster changes which subagent_type values are available for routing, and description changes can introduce or resolve routing confusion; for perm operations confirm both settings.json and permissions-guide.md are updated

End your response with a ## Confidence block per CLAUDE.md output standards.

• Atomic updates: always write-before-delete to prevent data loss on interruption
• Perm operations are dual-file: add perm and remove perm MUST update both settings.json and permissions-guide.md — they must stay in sync. Never update one without the other.
• settings.json format: Python json.load/json.dump with indent=2 is the safe editing path — avoids fragile sed/awk surgery on JSON. The output format (2-space indent, no trailing commas) matches the existing file style.
• No auto-edit for agent/skill operations: this skill only mentions when new permissions might be needed for create/update/delete — it does not mutate settings.json for those operations
• README.md tables: Step 7 updates the agent/skill tables in the project README.md — keep the row format consistent with existing rows
• Color pool: the AVAILABLE_COLORS list provides unused colors for new agents; if exhausted, reuse colors with a note
• Cross-ref grep is broad: searches bare kebab-case names across all markdown files — catches backtick references, prose mentions, spawn directives, and inventory lists
• MEMORY.md inventory: always regenerated from disk (ls), never manually calculated — this prevents drift
• Follow-up chains:
  • After any create/update/delete → /audit to verify config integrity, then /sync apply to propagate
  • After creating a new agent/skill → /review to validate generated content quality; for testing whether skill trigger descriptions fire correctly (trigger accuracy, A/B description testing), check the Anthropic skills repository for a skill-creator tool
  • After updating agent instructions (especially \<antipatterns_to_flag>) → /calibrate <agent> to measure whether recall and confidence calibration improved
  • After any agent create/update/delete → /calibrate routing fast to confirm routing accuracy is unaffected; a new agent changes the roster and its description must disambiguate from neighbors; deletion removes a valid routing target; rename changes the subagent_type identifier
  • After add perm/remove perm → /sync apply to propagate updated settings.json and permissions-guide.md to ~/.claude/
  • Recommended sequence for agent operations: /manage <op> → /audit → /calibrate routing fast → /sync apply