martinffx/atelier/spec-plan

Spec Plan

Write a proportional plan so clear that any engineer can follow it. The selected planning mode determines whether the plan stays in the conversation or becomes a persisted structured artifact. This skill does not write code or start implementation.

atelier-orchestrator owns automatic classification and the human may override it. When this skill is directly invoked without a selected mode, use Spec-backed Plan. If Inline planning reveals substantial design or coordination needs, ask the human whether to switch to a Spec-backed Plan before presenting the plan.

Terminal boundary

This skill produces exactly one plan output and then stops:

• Inline mode presents the plan in conversation and changes no repository files.
• Spec-backed mode creates or updates only plan.json from an approved design.md.

Planning drafts and annotation cycles stay in conversation. Do not modify design.md, persist a Markdown plan draft, create tracker entries, invoke another workflow skill, or edit implementation files. These rules still apply when the human asks to plan and implement in one request. Implementation requires a later, explicit request after this skill has finished.

Planning Rules

• Organize tasks around required behavior, not one task per layer.
• Record a short list of behavior the change must preserve.
• For each task, identify existing code to reuse, modify, or delete.
• Map every new abstraction to a present requirement and its current consumers.
• Introduce shared infrastructure only when at least two current consumers demonstrate the

same need.

• Group tests by changed contract and active boundary. Do not repeat the CRUD matrix across

layers.

• Plan tests at each module's interface, not past it (see oracle-codebase-design).

Proportionality Gate

Before presenting any plan, compare its size and concepts with the requested behavior. If a bounded migration or refactor introduces shared infrastructure, unrelated behavior, or a plan substantially larger than the behavior being changed, stop and simplify it.

Challenge Gate

Before presenting a plan, dispatch two read-only subagents in parallel with the draft plan and its design context. Select each agent explicitly (see Delegation in atelier-orchestrator):

1. Design review: check architectural fit, layer and data-model boundaries, API contracts,

and whether task ordering respects the dependencies.

1. Critical challenge: challenge assumptions, look for missing cases, scope creep, and

unproven requirements, and propose simpler alternatives.

Each returns at most five evidence-backed findings. Fold in the valid ones, drop the rest with a reason, and present the plan with a short Challenged note listing what changed.

Run the gate on every Spec-backed Plan. Run it on an Inline Plan only when the change touches a data model, an API contract, or more than one module or layer; skip it for small bounded edits. Run it once per draft; do not re-run it for the human's annotation revisions unless the plan's structure changes.

Outputs

Inline Plan (when explicitly selected)

No repository artifact and no task tracker entry. Read references/plan_template.md and present the plan in conversation using its structure. Omit optional subsections rather than rendering empty headings.

Inline Plan Workflow

Read enough of the codebase to identify the current behavior, boundaries, affected files, and concrete validation. Keep the plan proportional to implementation risk. Fill the template with confirmed, file-and-symbol-level details, including cross-file wiring or ordering constraints where they matter. Apply the Planning Rules, Proportionality Gate, and Challenge Gate before presenting it.

Do not manufacture phases, task IDs, dependency graphs, acceptance matrices, design.md, plan.json, tracker entries, or harness todos. The conversation is the plan artifact.

Tell the human: "Inline Plan ready for review."

STOP. Wait for human review.

If the human requests any adjustment, apply it to the working plan and re-present the entire updated plan using references/plan_template.md. Include unchanged sections so the human reviews one coherent plan. Never respond with only the changed text, an affected section, a summary, or an acknowledgement. A one-word correction is still a plan revision: incorporate it and present the complete plan again.

Every revision invalidates prior approval. End each revised plan with "Inline Plan ready for review." and stop until the human explicitly approves that complete version. Do not implement until the current plan is approved.

When the human approves the Inline Plan, acknowledge that the plan is approved and stop. Do not implement it in the same invocation. A later explicit implementation request may execute the approved conversational plan directly using the Inline execution safeguards in atelier-orchestrator; it does not use spec-implement, spec-finish, code-subagents, or task tracking.

Spec-backed Plan Artifacts

docs/specs/YYYY-MM-DD-<feature-name>/
├── design.md  ← From spec-brainstorm (approved)
└── plan.json  ← This skill's output

The plan starts as a conversational draft for human annotation, then gets converted to structured plan.json when approved.

plan.json Schema

{
  "feature": "user-authentication",
  "spec": "docs/specs/2026-03-08-user-auth/design.md",
  "goal": "Add email/password authentication with session management",
  "preserved_behavior": [
    "Existing sessions remain valid"
  ],
  "phases": [
    {
      "id": "P1",
      "name": "Authenticate with email and password",
      "tasks": [
        {
          "id": "T1",
          "name": "Accept valid credentials and reject invalid credentials",
          "depends_on": [],
          "inputs": [
            "User schema from design.md",
            "Validation rules (email format, password strength)"
          ],
          "description": "Create UserEntity with email and password fields. Implement validation using a Result type. Password must be hashed, never stored plaintext.",
          "files": {
            "reuse": ["src/auth/password.ts"],
            "create": ["src/entities/user.ts", "tests/entities/user.test.ts"],
            "modify": [],
            "delete": []
          },
          "new_abstractions": [
            {
              "name": "UserEntity",
              "requirement": "Validate email/password credentials",
              "consumers": ["POST /sessions"]
            }
          ],
          "validation": {
            "tests": [
              "rejects empty email",
              "rejects invalid email format",
              "rejects weak password",
              "accepts valid user input"
            ],
            "acceptance": [
              "All validation tests pass",
              "User.fromRequest returns Result<User>",
              "No direct throws — all errors via Result"
            ]
          }
        }
      ]
    }
  ]
}

Top-level fields

Phase fields

Task fields

Spec-backed Step 1: Write the Plan Draft

Read the approved design.md, apply the Challenge Gate to the draft, then present the plan draft in conversation. Do not write the draft into design.md or create a separate draft file.

Plan quality

Write assuming the implementer:

• Knows the language and framework
• Doesn't know this codebase's specific patterns
• Needs clear boundaries and validation criteria
• Will take the path of least resistance if the plan is vague

Task structure

Each task should be self-contained and include:

• Inputs: What you need to know or have before starting (from spec, existing code)
• Description: What to build, key design decisions, constraints
• Files: Exact existing code to reuse, modify, or delete, and files to create
• New abstractions: Present requirement and current consumers for each; use an empty list

when the task adds none

• Validation: Tests grouped by changed contract and active boundary, plus acceptance criteria

No exact code snippets. No implementation details. The task says WHAT and HOW TO VERIFY, not HOW to write the code.

Task ordering

Order tasks by required behavior and real dependencies. Keep changes across layers in one task when they implement and verify one contract; do not create one task per layer.

Task size

Each task should take 15-60 minutes. If larger, decompose into smaller tasks.

Tell the human: "Plan draft is ready for review."

STOP. Wait for human review.

Spec-backed Step 2: The Annotation Cycle

The human annotates the plan draft directly — adding corrections, rejections, domain knowledge, business constraints, or "remove this entirely."

You write plan → Human adds inline notes → You address all notes → Repeat 1-6x

When the human says "I added notes":

1. Re-read the entire document
2. Address every single note
3. Update the plan
4. Do not create tasks. Do not implement.

The "don't implement yet" guard is sacred. The plan is not ready until the human explicitly approves it.

Steering patterns

Spec-backed Step 3: Create Structured Tasks

When the human approves — "looks good", "approved", "create tasks" — convert the plan into plan.json.

What to do

1. Convert the annotated plan draft into structured plan.json
2. Each task maps to a unit with inputs, description, files, and validation
3. Dependencies between tasks are captured in depends_on fields
4. Do not create tracker entries or modify any file other than plan.json

Verification