bmad-autonomous-planning

name

bmad-autonomous-planning

description

Creates the full BMAD planning artifact chain (research, brief, PRD, architecture, epics/stories, readiness) autonomously from a short task description or GitHub issue, delegating each planning phase to a focused subagent and orchestrating the handoffs without human interaction. Use when /sdlc-plan runs stage 2 of the SDLC loop, or when the user wants BMAD-style planning from a short prompt without walking through interactive menus.

BMAD Autonomous Planning

Profile keys consumed

• project.name, project.repo
• php.version
• framework.name, framework.api_platform
• persistence.mapper, persistence.engine
• architecture.source_root, architecture.bounded_contexts, architecture.shared_context
• quality.phpinsights.complexity, quality.infection_msi

This skill invokes no make.* targets: the planning flow runs entirely in the current AI session. Quality thresholds appear in planning artifacts only as the profile's quality.* values, which are raise-only floors (shipped defaults: complexity 94 via quality.phpinsights.complexity, MSI 100 via quality.infection_msi) — a plan may tighten them, never relax them.

Non-Negotiable Rules

• Run the planning flow in the current AI session. Do not depend on repo-local bash wrappers, Make targets, or other launcher automation.
• Use BMAD as the primary process surface: start from _bmad/COMMANDS.md (and the repository's local bmalph skill wrapper when one exists), frame every subagent around a BMAD command name from that catalog, and only descend into the specific workflow or agent files required by that command.
• Spawn one focused subagent per BMAD planning stage when subagents are available. Do not overload a single subagent with the whole planning flow.
• Spawn planning and validation subagents with the strongest available model in the current platform, at the highest reasoning effort the platform supports. Do not downshift stage-owning subagents to a smaller or faster model.
• The main agent is the user surrogate. If a BMAD workflow asks for approval, choices, or clarification, decide on the user's behalf, continue, and record the decision inline in the artifact as > Assumption: <decision and rationale> instead of blocking.
• Do not implement production code during this skill. Produce specs only.
• Plans must preserve the root-cause culture: never propose lowering a quality.* threshold, adding suppression annotations, or editing deptrac.yaml — plan the code change that meets the bar instead.

Inputs

Expect the caller to provide:

• a short task description, or a GitHub issue (URL) whose title, problem statement, and acceptance criteria seed the chain — resolve it with gh issue view <url> --json url,number,title,body
• an optional bundle id or target bundle directory
• an optional validation round limit from 1 to 3
• optional GitHub issue or specs-only PR output requirements

Bundle directory resolution:

• When invoked from /sdlc-plan (or any caller that supplies an issue), use specs/<slug>/, where <slug> is the kebab-case issue title prefixed with the issue number (e.g. specs/42-currency-crud/).
• Otherwise derive one under the BMAD-configured planning artifacts path, for example <planning_artifacts>/autonomous/<timestamp>-<task-slug>.

Output Bundle

Create a planning bundle with at least these artifacts, in chain order:

• research.md
• brief.md (plus brief-distillate.md when it adds value)
• prd.md
• architecture.md
• epics-stories.md
• readiness.md — implementation-readiness verdict, an explicit PASS/FAIL with named findings
• run-summary.md

Cross-references must stay consistent: each artifact links its predecessors, and no artifact contradicts an upstream decision.

run-summary.md must also contain:

• the chosen bundle directory
• a Subagent Execution Log section listing the phase, BMAD command, and artifact owned by each subagent
• the validation rounds used per artifact
• open questions, warnings, blockers, and the recommended next step

The final assistant response should summarize:

• bundle directory (so a calling command such as /sdlc-plan can emit its SPECS_DIR: handle)
• artifact paths
• validation rounds used
• remaining open questions or warnings
• GitHub issue/PR status when requested

Required Sources

Load only the minimum sources required for the current stage:

1. _bmad/COMMANDS.md
2. The repository's local bmalph skill wrapper, when one exists (check the repo's agent-skills directories for a bmad/bmalph skill)
3. The resolved BMAD config file:
   • _bmad/config.yaml when present
   • otherwise _bmad/bmm/config.yaml
   • if both exist, treat _bmad/bmm/config.yaml as optional upstream context
4. Only the backing agent and workflow files required to satisfy the BMAD commands selected for this run:
   • analyst
     • _bmad/bmm/agents/analyst.agent.yaml
   • create-brief
     • _bmad/bmm/workflows/1-analysis/bmad-create-product-brief/workflow.md
     • _bmad/bmm/workflows/1-analysis/bmad-create-product-brief/steps/step-01-init.md
   • create-prd
     • _bmad/bmm/agents/pm.agent.yaml
     • _bmad/core/tasks/bmad-create-prd/workflow.md
     • _bmad/core/tasks/bmad-create-prd/steps-c/step-01-init.md
   • create-architecture
     • _bmad/bmm/agents/architect.agent.yaml
     • _bmad/bmm/workflows/3-solutioning/bmad-create-architecture/workflow.md
     • _bmad/bmm/workflows/3-solutioning/bmad-create-architecture/steps/step-01-init.md
   • create-epics-stories
     • _bmad/bmm/workflows/3-solutioning/bmad-create-epics-and-stories/workflow.md
     • _bmad/bmm/workflows/3-solutioning/bmad-create-epics-and-stories/steps/step-01-validate-prerequisites.md
   • implementation-readiness
     • _bmad/bmm/workflows/3-solutioning/bmad-check-implementation-readiness/workflow.md
     • _bmad/bmm/workflows/3-solutioning/bmad-check-implementation-readiness/steps/step-01-document-discovery.md
5. Repository guidance that constrains implementation: the project profile (.claude/php-sdlc.yml), agent guidance files (CLAUDE.md, AGENTS.md), and the repository's architecture/onboarding/developer docs where they exist

Never bulk-scan the whole repository when a narrow set of files will do.

Stage-to-Command Map

Use these BMAD commands as the default stage entrypoints for autonomous planning subagents:

• research: analyst
• product brief: create-brief
• PRD: create-prd
• architecture: create-architecture
• epics and stories: create-epics-stories
• implementation readiness: implementation-readiness

When a validation round needs another subagent pass, prefer the matching validation command when it exists, for example validate-brief, validate-prd, validate-architecture, or validate-epics-stories.

Main-Agent Responsibilities

The main agent owns orchestration and artifact quality. It must:

1. Resolve the bundle path and initialize the planning run.
2. Map each stage to a concrete BMAD command before spawning a subagent.
3. Decide which repository files each subagent needs.
4. Spawn the stage subagent with the strongest available model at the highest supported reasoning effort, and only the minimum context required.
5. Review the returned draft before moving to the next stage.
6. Answer workflow questions on behalf of the user, recording each as an inline > Assumption:.
7. Decide whether another validation round is necessary.
8. Maintain continuity across stages so the next subagent sees the correct upstream artifact set.
9. Update the Subagent Execution Log in run-summary.md after every phase.

Subagent Contract

For every BMAD stage, the main agent should hand the subagent:

• the specific BMAD command(s) from _bmad/COMMANDS.md it must execute
• the required runtime override: the strongest available model in the current platform, at the highest reasoning effort it supports
• only the backing workflow or agent files needed to fulfill those commands
• the current task framing
• only the upstream artifacts required for that stage
• only the repository files needed for evidence
• a clear output contract:
  • artifact draft content
  • key assumptions made
  • unresolved questions
  • validation findings or concerns

Every subagent should return a draft plus findings, not a request to pause for a human.

Do not hand a subagent only raw workflow-file paths without naming the BMAD command it is following.

Workflow

1. Preflight

• Resolve the BMAD config and planning artifacts directory.
• Create the bundle directory if needed.
• Read the local bmalph skill wrapper (when present) and _bmad/COMMANDS.md, then map the BMAD stage commands relevant to this planning run.
• Infer the most likely feature-area paths from the task description and the profile's architecture.bounded_contexts before any broad discovery.
• Write a short task framing section into run-summary.md.

2. Research Stage

Spawn a research subagent through the analyst BMAD command.

The research subagent should:

• inspect the most relevant docs and code paths
• summarize current-state behavior and constraints
• identify implementation risks and likely surface area
• return a draft for research.md

The main agent then reviews the result, resolves open choices, and finalizes research.md.

3. Product Brief Stage

Spawn a brief subagent through the create-brief BMAD command.

Inputs:

• task description (or resolved issue)
• research.md
• only the command entry and backing files needed for create-brief

Outputs:

• draft brief.md
• optional brief-distillate.md
• explicit gaps, risks, and questions

The main agent must review the draft before moving on.

4. PRD Stage

Spawn a PRD subagent through the create-prd BMAD command.

Inputs:

• research.md
• brief.md
• brief-distillate.md when present

The PRD subagent should produce an implementation-ready but not code-level prd.md, with the issue's acceptance criteria traced into FRs. Quality NFRs must cite the profile's quality.* values as raise-only floors — never a lowered target. The main agent validates coverage, measurability, traceability, and completeness before proceeding.

5. Architecture Stage

Spawn an architecture subagent through the create-architecture BMAD command.

Inputs:

• research.md
• prd.md
• repository architecture guidance and relevant feature-area code

The architecture must fit the repository's actual stack and patterns as declared by the profile: php.version, framework.name (and framework.api_platform when set), persistence.mapper / persistence.engine, and the layered/bounded-context layout under architecture.source_root with architecture.bounded_contexts and architecture.shared_context.

# Reference-service fit statement derived from the profile:
# Symfony + API Platform service on doctrine-odm/mongodb, contexts
# [User, OAuth] plus shared kernel "Shared" under src/ — the plan must
# slot new components into those contexts, not invent new top-level dirs.

The main agent validates compatibility and implementation readiness before moving on. If the PRD is not strong enough, do not improvise; send the flow back to PRD refinement first.

6. Epics and Stories Stage

Spawn an epics/stories subagent through the create-epics-stories BMAD command.

Inputs:

• prd.md
• architecture.md
• relevant constraints from research.md

The subagent should produce forward-safe epics and actionable stories in epics-stories.md, each story marked independent or dependent (the parallel-dispatch input for the implementation stage). The main agent separately reviews story quality, dependency order, and acceptance-criteria coverage.

7. Cross-Artifact Readiness Stage

Spawn a readiness subagent through the implementation-readiness BMAD command.

Inputs:

• brief.md
• prd.md
• architecture.md
• epics-stories.md

This subagent should identify gaps, inconsistencies, and unresolved planning risks in readiness.md, ending in an explicit PASS/FAIL verdict with named findings (the artifacts each finding implicates). The main agent finalizes the readiness assessment and updates run-summary.md. On FAIL, correct the named artifacts and re-run readiness — the calling command bounds this loop.

Validation Loop

Use 1 to 3 validation rounds per artifact.

For each artifact, the main agent may:

• accept the draft
• revise it directly
• spawn a reviewer subagent for another pass, preferably using the matching BMAD validation command when one exists and keeping the strongest-available-model override

Stop early when only minor or repetitive issues remain. Do not loop endlessly.

Decision Policy for Interactive BMAD Gates

If a BMAD wrapper or workflow expects user input:

• continue without asking the human — zero interactive prompts anywhere in the chain; never use AskUserQuestion, never wait for confirmation
• choose the best option based on task intent and repository evidence, and record it inline as > Assumption: <decision and rationale>
• prefer another review round when uncertainty is material
• record unresolved concerns in run-summary.md and readiness.md

If a phase is blocked by a genuinely missing prerequisite, stop only that phase, record the blocker explicitly, and do not fabricate the missing input.

Use the BMAD menu concepts as policy, not as a hard stop:

• deeper review when the artifact is still weak
• extra perspective when another subagent is likely to add signal
• continue when the artifact is ready enough for the next stage

GitHub Output

Only create a GitHub issue or specs-only PR when the user explicitly asks.

When requested:

• finish the planning bundle first
• create GitHub side effects only after the artifacts are stable, against project.repo
• prefer GitHub app tools when available
• fall back to gh only when necessary
• record failures as warnings instead of discarding the planning bundle

Related Skills

• bmad-fr-nfr-review-gate — the post-implementation counterpart: verifies the implemented change against the specs this skill produces
• quality-standards — the protected quality.* thresholds that PRD and architecture NFRs must restate
• implementing-ddd-architecture — the layer and pattern rules architecture plans must fit