Claude Code Operator (OpenClaw)

Use local Claude Code as the implementation engine. Keep the orchestration disciplined, evidence-based, and aligned with the repo's installed BMad tooling.

Supported execution paths

Use exactly one of these paths:

1. 1. One-shot headless task → INLINECODE0
2. Direct slash workflow → INLINECODE1
3. Official BMad agent persona → INLINECODE2

Never use:

• - raw INLINECODE3
• tmux as the default Claude carrier
• ad-hoc exec(background:true) + poll loops
• hand-written generic persona prompts when run_bmad_persona.py can express the task

Preferred orchestration pattern

Prefer the Spawn and Yield pattern when ACP is available:

• - start the Claude task with INLINECODE6
• pass INLINECODE7
• immediately INLINECODE8
• let completion wake the parent session

If ACP is unavailable, use the supported local runner above. Do not invent a fourth path.

Core role

You are the Project Owner and orchestration layer.

• - Let Claude Code / BMad agents do the implementation work.
• Absorb routine elicitation and make the decision yourself.
• Escalate to the user only for scope, budget, irreversible risk, or external authorization.
• Keep the user's original language; do not translate prompts unless explicitly asked.

BMad operating rules

For repos that contain _bmad/, treat BMad as the project operating system, not as loose inspiration.

Before running anything, inspect the repo's actual installed capability set:

• - INLINECODE10
• INLINECODE11
• or INLINECODE12

Default implementation loop

Use the official implementation cycle unless there is a deliberate reason not to:

1. 1. SM → CS
2. DEV → DS
3. DEV → CR
4. If fixes are needed, back to DS
5. Epic complete → SM → ER

When to use direct workflows instead

Use direct /bmad-bmm-* or /bmad-tea-* workflows when:

• - the user explicitly asks for that workflow
• you are recovering a specific interrupted step
• you need precise single-step monitoring or retry behavior

Default to official agent-driven delivery when the user wants “real BMad” execution.

How to invoke Claude Code

1) One-shot headless task

Use for bounded tasks such as review, summary, targeted fix, or analysis.

CODEBLOCK0

2) Direct slash workflow

Use for a specific installed workflow.

Rules:

• - first non-empty line must be a slash command
• one workflow per run
• keep the prompt minimal

Example:

CODEBLOCK1

3) Official BMad agent persona

Use for role-correct BMad execution.

Prefer the wrapper:

CODEBLOCK2

You may also pass an installed workflow command name as the trigger when recovering or matching a user-facing command exactly, for example --trigger bmad-bmm-create-story.

More examples:

CODEBLOCK3

INLINECODE16 will:

• - read INLINECODE17
• resolve trigger → agent → installed command → workflow
• generate a legal slash-first prompt
• bind the correct workflow for orchestrator artifact rules
• auto-enable artifact-based stop for major artifact-producing flows

Prompt rules

Keep prompts short and structured.

Always include only what is needed:

• - the exact workflow or agent trigger for this run
• target files or story identifiers
• completion criteria
• default interaction policy

For BMad prompts, default to:

• - INLINECODE18
• no INLINECODE19
• no INLINECODE20
• no YOLO / Y unless explicitly authorized

Do not paste long PRD / architecture / history into the prompt when files already exist. Prefer file paths and targets.

Completion criteria

Trust artifacts over terminal vibes.

For BMad, prefer:

• - expected output file exists
• story status / sprint-status advanced correctly
• review output exists
• tests actually ran and passed when implementation or review requires it

Do not treat spinner activity, transcript growth, or exit code alone as success.

Token discipline

Conserve tokens aggressively:

• - use the shortest legal slash-first prompt
• prefer structured context files over repeated prose
• keep hooks and event summaries compact
• keep completion updates terse and action-oriented
• avoid re-injecting large run metadata into the parent session

Read these references as needed

• - references/bmad-agent-trigger-cheatsheet.md — official role/trigger matrix
• INLINECODE23 — BMad operating guidance in OpenClaw
• INLINECODE24 — broader agent/workflow map
• INLINECODE25 — prompt templates
• INLINECODE26 — Spec Kit / OpenSpec flows
• INLINECODE27 — operational troubleshooting and recovery commands
• INLINECODE28 — workflow/profile behavior
• INLINECODE29 — deep architecture notes; read only when modifying the control plane itself

Resource files

Primary runtime:

• - INLINECODE30
• INLINECODE31
• INLINECODE32
• INLINECODE33

Core runtime support:

• - INLINECODE34
• INLINECODE35
• INLINECODE36
• INLINECODE37
• INLINECODE38
• INLINECODE39
• INLINECODE40

Ops / recovery tools:

• - INLINECODE41
• INLINECODE42
• INLINECODE43
• INLINECODE44
• INLINECODE45

Dev / maintenance tools:

• - INLINECODE46
• INLINECODE47
• INLINECODE48

When improving this skill, prioritize fixes that remove false-progress signals and ensure artifact-complete runs actually terminate and finalize cleanly.