Orchestrate Initiative Planning

Orchestrate Initiative Planning

Purpose

Use this skill when the user wants one agent to coordinate the full planning pipeline:

1. Product definition
2. Technical architecture
3. Execution planning

This skill is an orchestrator. It should launch sub-agents, manage handoffs, answer sub-agent questions when possible, and stop before implementation.

Workspace Isolation

All work MUST happen in a git worktree.

Before doing anything else — before reading source docs, before launching sub-agents — enter a worktree:

1. Use EnterWorktree with a name derived from the initiative (e.g., plan-).
2. All subsequent file creation and sub-agent work happens inside the worktree.
3. Do not write any planning artifacts to the main working tree.

If EnterWorktree fails (e.g., already in a worktree), halt and ask the user how to proceed.

Artifact Contract

The orchestrated workflow should produce:

1. Product artifacts in tasks//

• VISION.md (large initiatives only)
• PRD..md

1. Top-level VISION.md update — high-level entry capturing the initiative's intent (large initiatives only)
2. Architecture docs in docs/architecture/ — organized by system or topic, not by initiative
3. Execution artifacts in tasks//

• EPIC.md
• task files

Core Rule

Do not implement product code in this workflow.

plan-eng-tasks is planning-only — it produces task files but does not implement them. When execution is needed after planning, use exec-eng-tasks to dispatch sub-agents for implementation.

In this orchestrated workflow, plan-eng-tasks produces:

• updated EPIC.md
• task files
• any required architecture doc clarifications

It does not dispatch implementation sub-agents or write code. If the user wants to proceed to execution after the planning PR is merged, invoke exec-eng-tasks separately.

Stage Workflow

Stage 1: CEO Product Review

Launch a sub-agent that applies plan-requirement-doc.

Its job:

• review and improve the source PRD or planning doc
• refine ICP, problem, value proposition, positioning, metrics, and launch narrative
• create or update:
• tasks//VISION.md (large initiatives)
• tasks//PRD..md
• top-level VISION.md with a high-level entry for the initiative (large initiatives)

Stage 1 is complete only when the required product artifacts exist and are updated.

Stage 2: CTO Technical Review

After Stage 1 completes, launch a new sub-agent that applies plan-tech-spec.

Its inputs:

• tasks//VISION.md (if exists)
• relevant tasks//PRD..md

Its job:

• review the initiative from a technical perspective
• update or create architecture docs in docs/architecture/ (organized by system or topic, not by initiative)
• create one or more EPICs in tasks//EPIC.md

When architecture drafting is needed, instruct the sub-agent to use draft-technical-architecture.

Stage 2 is complete only when architecture docs and initial EPICs exist.

Stage 3: Engineering Task Planning

After Stage 2 completes, launch a separate sub-agent for each newly created EPIC.

Each sub-agent applies plan-eng-tasks.

Scope is already locked. The orchestrator confirmed scope in Stages 1-2. Instruct each sub-agent to skip Step 0 (Scope Challenge) and begin directly at the readiness sections. Include this directive in the sub-agent prompt: SKIP_STEP_0=true — scope was locked by the orchestrator during initiative planning. Do not re-challenge scope.

Each sub-agent consumes:

• the relevant docs in docs/architecture/
• the relevant tasks//EPIC.md

Its job:

• review the EPIC for execution readiness
• update the EPIC if needed
• create or refine task files under tasks//
• update architecture docs only when execution-planning discoveries require clarification

Stop condition:

• no code changes
• no implementation sub-agents
• no execution approval step
• planning artifacts only

Run these EPIC-planning sub-agents in parallel only when the EPICs are independent and will not create conflicting task boundaries.

Stage 4: Push and Open PR

After all planning stages complete, deliver the artifacts:

1. Stage all new and modified files in the worktree.
2. Create a commit with a clear message summarizing the planning artifacts produced (e.g., plan(): add vision, architecture, and task artifacts).
3. Push the worktree branch to the remote with git push -u origin .
4. Open a pull request using gh pr create:

• Title: plan(): initiative planning artifacts
• Body must include:
• Summary of artifacts created (vision docs, architecture docs, EPICs, task files)
• List of unresolved decisions or remaining risks
• Note that this is a planning-only PR — no product code changes

1. Return the PR URL to the user.

If the push or PR creation fails, report the error and the branch name so the user can recover manually.

Orchestrator Behavior

Answering Questions

You should answer sub-agent questions directly when possible.

Escalate to the user only when:

• the sub-agent is genuinely blocked
• the answer cannot be resolved from existing artifacts or repository context
• the decision is strategic or high-impact enough to require user input

When escalating, summarize:

• what stage is blocked
• what the sub-agent is deciding
• your recommended option

Handoff Discipline

Before starting each stage:

1. Verify the previous stage's required artifacts exist.
2. Summarize the prior stage in 3-5 bullets.
3. Pass those artifacts explicitly into the next sub-agent's prompt.

Traceability

Maintain this chain:

• PRD requirement -> Architecture decision -> EPIC -> task file

Call out gaps when:

• an EPIC has no PRD anchor
• a task has no EPIC anchor
• architecture work is not justified by product artifacts

Kickoff Template

Use this structure when the user asks to run the full workflow:

Act as the orchestration agent for a 4-stage planning workflow. Use the installed
skills to move from product definition -> technical architecture -> execution
planning -> delivery.

STAGE 0: WORKTREE SETUP
1. Use EnterWorktree with name `plan-` to create an isolated worktree.
2. All subsequent work happens inside the worktree.
3. If worktree creation fails, stop and ask me how to proceed.

STAGE 1: CEO PRODUCT REVIEW
1. Launch a sub-agent and have it apply `/plan-requirement-doc` to review and improve
   .
2. Produce:
   - `tasks//VISION.md` (large initiatives)
   - `tasks//PRD..md`
   - Update top-level `VISION.md` with high-level intent (large initiatives)
3. Answer the sub-agent's questions when possible.
4. Only escalate to me when genuinely blocked.
5. Do not move to Stage 2 until the product artifacts are updated.

STAGE 2: CTO TECHNICAL REVIEW
1. Launch a new sub-agent and have it apply `/plan-tech-spec`.
2. Inputs:
   - `tasks//VISION.md` (if exists)
   - `tasks//PRD..md`
3. Produce:
   - updated or new docs in `docs/architecture/` (by system/topic, not initiative)
   - one or more EPICs in `tasks//EPIC.md`
4. Use `/draft-technical-architecture` when helpful.
5. Answer the sub-agent's questions when possible.
6. Do not move to Stage 3 until the architecture docs and EPICs exist.

STAGE 3: ENGINEERING TASK PLANNING
1. Launch a separate sub-agent for each EPIC.
2. Have each sub-agent apply `/plan-eng-tasks`.
3. Tell each sub-agent: `SKIP_STEP_0=true — scope was locked by the orchestrator. Do not re-challenge scope.`
4. Inputs:
   - relevant docs in `docs/architecture/`
   - `tasks//EPIC.md`
5. Produce:
   - updated EPICs
   - task files
   - architecture doc clarifications only if needed
6. IMPORTANT: Do not implement anything.
7. Stop after planning artifacts are complete.

STAGE 4: PUSH AND OPEN PR
1. Stage and commit all planning artifacts.
2. Push the worktree branch to origin.
3. Open a PR with `gh pr create`.
4. Include in the PR body: artifacts created, unresolved decisions, and a note
   that this is a planning-only PR with no product code.
5. Return the PR URL to me.

ORCHESTRATION RULES
- Maintain path discipline:
  - Product artifacts in `tasks//` (VISION.md, PRDs)
  - Architecture docs in `docs/architecture/` (by system/topic)
  - Execution artifacts in `tasks//` (EPIC, task files)
- Preserve traceability from product -> architecture -> execution.
- Answer sub-agent questions directly when possible.
- Escalate only when genuinely blocked.
- Run Stage 3 in parallel only when EPICs are independent.
- End with a concise summary of artifacts created, unresolved decisions, EPICs,
  task files, and remaining risks.

Final Output

At the end of the orchestrated workflow, return:

1. PR URL
2. Branch name
3. Artifacts created or updated
4. EPICs produced
5. Task files produced
6. Unresolved decisions
7. Remaining risks or gaps

Source & license

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

• Author: dremonkey
• Source: dremonkey/agents-and-skills
• License: MIT

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