Mhlavac Claude Marketplace Agent Create

agent:create

Scaffold a Claude Code subagent that has durable memory — not just a system prompt, but a small filesystem the agent reads at boot and writes back to at the end of each session.

The skill creates three things in one shot:

Files are templates. They include placeholders, hints, and a few sensible defaults (e.g. PRIVACY.md as the first boot step, a self-curation block in the definition, a tasks/open-tasks.md skeleton). The user customizes after the scaffold runs.

When this earns its place

Use it when the user is about to create — or wishes they had created — an agent that should:

• Have a persistent identity across sessions, not a one-shot prompt
• Remember decisions, preferences, and "what works" without being told each time
• Carry per-topic state (per-person profiles, per-project notes, etc.)
• Curate its own memory after each session (append to what-works.md, log a session, mark a task done)

Skip it for:

• One-off slash commands (those go in commands/, not agents/)
• Skills that wrap a tool — those don't need their own memory store
• "Just give me a system prompt" — that's a one-file edit, not three-tier scaffolding

The flow

1. Collect the minimum input

Before running the scaffolder, confirm with the user:

1. Name — kebab-case, used in all three tiers (e.g. code-reviewer, release-notes-writer, data-scout). The agent file will be <name>.md; the deep store will live at .agents/<name>/.
2. Role / one-line purpose — a single sentence the user would say if asked "what does this agent do for me?" Used in the description-trigger and in the identity blocks of all three tiers.
3. Target directory — where the three tiers should be created. Default is the current working directory. For a vault-style setup, pass the vault root (the scaffolder will create .claude/agents/, .claude/agent-memory/, and .agents/ under it).

Optional fields (the scaffolder accepts but leaves blank if not provided):

• --anchor — a short anchor phrase the agent returns to ("Boring is the feature.", "Signal, not noise.")
• --voice — one line of voice notes ("calm, dry, no exclamations, no emojis")
• --triggers — comma-separated keywords for the description ("X, Y, Z")
• --anti-triggers — comma-separated "do NOT trigger for" topics
• --force — overwrite existing files (default: refuse if any of the three tiers already exists)

2. Run the scaffolder

# $CLAUDE_PLUGIN_ROOT is set by Claude Code when the plugin is installed
python3 "$CLAUDE_PLUGIN_ROOT/skills/agent:create/scripts/create_agent.py" \
  --name <agent-name> \
  --role "<one-line purpose>" \
  --target <vault-or-project-root> \
  [--anchor "<phrase>"] \
  [--voice "<one line>"] \
  [--triggers "kw1, kw2, kw3"] \
  [--anti-triggers "topic1, topic2"]

The script prints the absolute paths of every file it created and exits 0 on success, non-zero with a clear message if a tier already exists (use --force to overwrite, or pick a different name).

3. Review and customize

After the scaffold lands, walk the user through the three files that matter most:

1. **.claude/agents/<name>.md** — fill in persona, boundaries, output formats. The template ships with section headings and <!-- TODO --> markers so it's obvious what's missing.
2. **.claude/agent-memory/<name>/MEMORY.md** — sanity-check the boot sequence. Add or remove file pointers based on what memory/ actually contains.
3. **.agents/<name>/memory/PRIVACY.md** — set the privacy posture before the agent is ever invoked. The template has prompts for what to redact, what to keep in buckets, and what to point-to-vault rather than duplicate.

4. Run a first audit

After customization, the user can run mh:agent:audit (the sister skill in this plugin) against the new agent to check that:

• All boot-sequence files referenced in MEMORY.md actually exist
• The description has both positive triggers and "do NOT trigger" rules
• PRIVACY.md, tasks/, and sessions/ are present
• Cross-tier consistency holds (name, identity one-liner)

This closes the loop: scaffold → customize → audit → iterate.

What the scaffolder creates

<target>/
├── .claude/
│   ├── agents/
│   │   └── <name>.md                 ← Tier 1 — definition + system prompt
│   └── agent-memory/
│       └── <name>/
│           └── MEMORY.md             ← Tier 2 — auto-injected boot index
└── .agents/
    └── <name>/
        ├── memory/
        │   ├── PRIVACY.md            ← always read first by the agent
        │   ├── INDEX.md              ← pointers into the deep store
        │   └── _shared/.gitkeep      ← framework / pattern files (load on demand)
        ├── tasks/
        │   └── open-tasks.md         ← agent's own backlog (Markdown checkboxes)
        ├── sessions/
        │   └── .gitkeep              ← session logs land here per ritual
        ├── cache/
        │   └── .gitkeep              ← ephemeral cache (rates, RSS poll state, …)
        └── output/
            └── .gitkeep              ← draft outputs before publish

Templates live at templates/ inside this skill. They are deliberately short: they give shape, not content.

Conventions baked into the templates

• PRIVACY.md is step 1 of every boot sequence. Always.
• MEMORY.md kept ≤80 lines. It pays its cost on every spawn — keep it lean. Detail goes to memory/INDEX.md and topic files.
• Boot sequence is a numbered list. No "read these files" prose — the order matters and should be explicit.
• Self-curation block in the definition. Every agent writes back at session end: what-works, decisions, tasks closed, MEMORY.md trim check.
• Subagents do NOT inherit the CLAUDE.md cascade is called out in the template — boot sequence should explicitly list any CLAUDE.md the agent depends on.

What the skill does NOT do

By design, this skill stays out of the user's domain decisions:

• No cadence integration — no daily/weekly/monthly assumptions, no embedding the agent in any periodic-routine flow. If the user wants that, they wire it after.
• No external-system handoffs — no MCP wiring, no queue conventions, no cron/scheduling. The scaffolded agent is fully self-contained.
• No persona invention — anchor phrase, voice, triggers, and tool boundaries are user input. Templates have placeholders, not opinions.
• No vault opinions — the scaffolder writes the three tiers wherever you point it. It does not touch existing CLAUDE.md, README, or git config.

Inputs / outputs

Inputs: name, role, target dir, optional anchor / voice / triggers / anti-triggers. Outputs: three tiers of files described above, absolute-path-listed on stdout. Exit codes: 0 on success; non-zero with a clear message on collision or invalid input.

See also

• mh:agent:audit — sister skill, checks the structure for completeness and cross-tier consistency.
• The plugin README at plugins/mh/README.md.

Source: mhlavac/claude-marketplace — distributed by TomeVault.