Deepworkplan Create

DeepWorkPlan — Create

Create a new Deep Work Plan through a smooth, unified flow: the developer provides information once, you generate a single refined draft, the developer reviews it, and you materialize the final plan under .dwp/plans/PLAN_{name}/.

Single-step change (vs legacy): DeepWorkPlan generates the refined draft directly. There is no separate raw-draft file. This replaces the legacy draft → refined-draft two-step (which wrote PLAN_{name}_draft.md and then PLAN_{name}_draft_refined.md). The only reviewable draft artifact is now .dwp/drafts/PLAN_{name}_draft_refined.md.

Philosophy

The goal is a delightful, smooth experience. The user provides information once; the system handles all intermediate steps (refined-draft creation, review, final plan generation) automatically.

Shared resources (read these)

• ../shared/context.sh — resolve repo root, branch, agent tool, and the .dwp/ output location (dwp_dir).
• ../shared/dwp-paths.md — the .dwp/plans/ + .dwp/drafts/ output convention.
• ../shared/adaptation.md — reasoning-over-copy-paste and the two repository archetypes (individual repo vs orchestrator hub).
• ../guide/GUIDE.md — the full methodology: plan README structure, task-file anatomy, mandatory final tasks, orchestrator (§13), team agents (§14).
• ../examples/CREATE_PLAN.md — prompt patterns.
• ../examples/PROMPTS_TEMPLATE.md — the PROMPTS.md template for each plan.

Parameter Reference

Name format: users type names in any format; you auto-convert to snake_case internally.

Modes

Guided Mode (default)

• Collects information from the user.
• Creates the refined draft → shows it for review.
• Asks for confirmation before creating the final plan.
• User can request adjustments before final generation.

Trust Mode (trust or auto)

• Collects information from the user.
• Creates the refined draft → final plan automatically (no intermediate confirmations).

Trust boundary (write scope)

allowed-tools includes write-capable Edit, Write, and Bash.

Writes: plan artifacts under the gitignored .dwp/ directory only — .dwp/drafts/ during drafting, .dwp/plans/PLAN_{name}/ for the materialized plan (README, task files, analysis outputs). "Trust mode" skips intermediate confirmations of plan content, not of the write boundary.

It MUST NOT: modify source files (that is execute's job), write outside .dwp/, read or include secrets in plan content, or materialize a plan whose tasks lack acceptance criteria and validation gates.

Unified Workflow

Step 0 — Parse Parameters & Determine Mode

0.1 Detect trust mode: if the LAST word is trust or auto, remove it and set trust_mode = true; otherwise false.

0.2 Detect special modes (check FIRST word):

• refined-draft → mode = "refined-draft-only", remaining text = plan name.
• from-refined-draft or from → mode = "from-refined-draft", remaining text = refined-draft file path.
• Otherwise → continue to input classification (0.3).

0.3 Classify remaining input:

Name auto-conversion to snake_case: lowercase; replace hyphens/spaces with _; strip everything but a-z0-9_. Examples: improve Feature X → improve_feature_x, add-stripe-payments → add_stripe_payments.

CRITICAL: when input is name-only, NEVER explore the codebase or research the topic before asking questions. The name only says what to CALL the plan, not what to DO. Go directly to Step 2.

Routing by mode:

• from-refined-draft → skip to Step 4.3.
• refined-draft-only → gather info, create only the refined draft, skip final plan creation.
• no input / name-only → Step 1, then Step 2.
• full-context → Step 1, then Step 3.

Step 1 — Quick Introduction

Show a brief intro matching the mode (guided vs trust): what will happen and that you'll create a refined plan, review it (guided) or proceed (trust), then generate the final executable plan.

Step 2 — Gather Information (Conversational)

Skip this entire step for full-context input — go straight to Step 3.

Collect, conversationally:

• 2.1 Plan name (skip if already extracted) — auto-convert to snake_case, add PLAN_ prefix internally.
• 2.2 Objective — one or two sentences.
• 2.3 Context — where the changes live, constraints/rules, tech notes.
• 2.4 Tasks — at least 2; if only 1, suggest breaking it down.
• 2.5 Guidelines (optional) — branch/commit format, coverage target, etc.

2.6 Orchestrator detection (automatic). After 2.3–2.4, auto-detect an orchestrator plan when: the work spans 2+ sub-repositories with independent feature work, or the user explicitly mentions child DWPs / orchestrator / "create plans in each repo". If detected, offer the choice between an orchestrator plan (creates child DWPs per repo, each following its own AGENTS.md) and a direct multi-project plan. See ../guide/GUIDE.md §13 and ../shared/adaptation.md (orchestrator-hub archetype).

2.7–2.9 (orchestrator only): ask for target repositories (note each repo root — for an orchestrator hub this is repositories/{repo}/), dependency order, and execution mode (Distributed / Sequential with Output Handoff / Sequential basic). Default to Distributed.

2.10 Team-agents detection (automatic — always runs, non-orchestrator plans). Always analyze whether 2+ tasks touch different files/modules with no data dependencies and would benefit from parallel execution. This is NOT opt-in.

• If parallelizable: in guided mode, inform the user (do not ask) and add the team-agents configuration; in trust mode, do it silently. Team-agents metadata is always additive and backward compatible (other agents ignore it).
• Auto-assign parallel groups (tasks with no cross-dependencies), teammate roles (derived from task content), and default model sonnet. Setup/integration and the three mandatory final tasks are always sequential.
• If not parallelizable: add nothing, mention nothing.

Step 2.11 — Parallel Research Phase (Claude Code only, automatic). Before drafting, if the plan spans 2+ repos or several independent modules and context isn't already provided, spawn research teammates (subagent_type: "Explore", one per repo/area) to read each AGENTS.md, identify relevant files, contracts, and validation commands, then synthesize their findings into enriched context for the refined draft. Skip for simple/single-module plans or full-context input. Fallback: research sequentially if team agents are unavailable. In trust mode, run silently.

Step 3 — Create the Refined Draft (Automatic, SINGLE STEP)

This is the single-step change. Produce the refined draft directly — one file, written straight to .dwp/drafts/. Do NOT create a separate raw PLAN_{name}_draft.md first.

Show a 2-step progress UI:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Creating your plan...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[1/2] Drafting refined plan...

3.1 Write the refined draft:

• Resolve dwp_dir via ../shared/context.sh; ensure <dwp_dir>/drafts/ exists.
• Compose a professional, complete plan prompt following ../examples/CREATE_PLAN.md: objective, context (enriched with Step 2.11 research if applicable), well-formed tasks, guidelines. Expand every section with full detail and clarity in one pass.
• Write only: .dwp/drafts/PLAN_{name}_draft_refined.md.

[1/2] Drafting refined plan... ✓
[2/2] Preparing for review...

Step 4 — Review & Create Final Plan

4.1 Guided Mode — Show for Review

Present a summary (objective, task count + list, location, constraints) and the path → .dwp/drafts/PLAN_{name}_draft_refined.md, then offer:

1. Looks good, create the final plan → Step 4.4.
2. Make adjustments → ask what to change, edit the refined draft in place, show the menu again.
3. Show the full refined draft → display it, then re-show the menu.
4. Stop here → completion message for the draft phase.

4.2 Trust Mode — Automatic Continue

Skip review; go directly to Step 4.4. Progress: [2/2] Creating final plan....

4.3 From-Refined-Draft Mode

• Validate the file exists in .dwp/drafts/. If not found, list available refined drafts and ask the user to select.
• Read it; extract plan name, objective, context, tasks, guidelines.
• Skip to Step 4.4.

4.4 Create Final Plan

Follow ../guide/GUIDE.md. Create:

1. Folder: .dwp/plans/PLAN_{name}/

2. README.md — Goal; Context; Plan Variables (optional); Global Guidelines; Task List with [ ] checkboxes + links (including the three mandatory final tasks); Execution Rules; Skills & Agents Used; Plan Status / Notes; Analysis Outputs (referencing EXECUTIVE_REPORT.md); Quick Reference to PROMPTS.md.

3. PROMPTS.md — from ../examples/PROMPTS_TEMPLATE.md, replacing {PLAN_NAME}.

4. PROGRESS.md — initial template (Task Summaries / Key Decisions / Important Values & Paths).

5. analysis_results/ — empty folder inside the plan directory.

6. User-defined task files — N.task_{title}.md, each with Context, Read Before Starting (optional), Goal, Instructions (with re-anchoring), Acceptance Criteria, Outputs (optional), Validation, Rollback (optional), Execution Checklist, Completion & Log. For any task that adds new core functionality or changes product behavior, bake the test discipline into it (../guide/GUIDE.md §5.3): its Acceptance Criteria must require automated test coverage for the new/changed behavior, and its Validation must run the repo's tests plus lint/type-check/format checks (not the build alone). Where related work is substantial, prefer a dedicated N.task_add_tests_for_{feature}.md task right after the implementation task. Likewise, for any task that touches auth, input handling, secrets/config, network surface, or dependencies, bake the security discipline into it (../guide/GUIDE.md §5.4); where the security-sensitive work is substantial, prefer a dedicated N.task_security_hardening_{feature}.md task placed after the implementation tasks and before the comprehensive-tests task, so findings are fixed before tests encode the behavior and become regression test cases rather than rework.

7. State layer (RECOMMENDED, ../spec/PLAN_STATE.md) — write manifest.json (plan identity: name, archetype, rigor tier, spec version, task count, creating agent — once, never edited after) and the initial state.json (every task pending, empty gates). Both atomically (write-temp-then-rename). REQUIRED when the plan will run unattended (../spec/AGENT_PROTOCOL.md §7.2) or the workspace has no git.

Proportional rigor check first (../spec/DWP_SPECIFICATION.md §11): before materializing anything, confirm the work actually warrants a plan. A trivial single-concern change is micro tier — say that a plan is disproportionate, offer to state goal + acceptance criteria + validation gate inline and just do it. Declare the chosen tier (standard or deep) and why in the refined draft.

Three mandatory final tasks (always):

• Security Review (MANDATORY, third-to-last, task N-2, {N-2}.task_security_review.md): reviews the plan's full accumulated diff for hardcoded secrets, injection risks, unsafe input handling, new attack surface, and auth/permission changes; audits dependencies the plan introduced (best-effort, with the ecosystem's audit tooling where available); verifies docs/SECURITY.md still reflects reality and updates it when the plan changed secrets handling, the auth model, or data boundaries; writes analysis_results/SECURITY_REVIEW.md even when clean. A critical finding blocks completion until fixed or explicitly accepted by the user (../spec/DWP_SPECIFICATION.md §6.1).

  Addon augmentation — ai-diff-reviewer (opt-in, only when installed). When the target repo has installed the ai-diff-reviewer addon — detected via .agents/skills/ai-diff-reviewer/ present + an extension file at one of the three recognized paths (in precedence order): .review/extension.md, .github/ai-diff-reviewer/extension.md, or the back-compat .github/ai-pr-reviewer/extension.md — the Security Review task template gains an ADDITIONAL post-existing-checks step: invoke the upstream skill's parent default flow ("Review my current branch" / /ai-diff-reviewer), capture the verdict, findings table, per-finding bodies, notes, and recommendation, and append them to analysis_results/SECURITY_REVIEW.md under a dedicated ## AI Diff Reviewer local review heading. The upstream skill's prompt.md is byte-identical to the CI Action's prompts/default.md at the same tag, so when the same repo also runs the CI Action (Flow B), the local review shares the same methodology and severity model via that prompt plus the extension file; CI round 2+ may surface a shorter finding set under Iteration-Aware Review (local stays a full pass — see addon SPEC §4.3). A critical finding follows the existing SR contract (blocks completion until fixed or explicitly accepted); warning / info findings are appended and reported but do not block. The augmentation is best-effort and conditional per the addon SPEC §7 (never-block rule): skip the local review pass (warn once, NEVER fail the task) only when the vendored skill is absent, detection fails (no extension file), or the local review invocation errors. Flow A needs no CI provider secret — do NOT treat an unset CURSOR_API_KEY (or other provider secret) as a reason to skip the local Security Review pass; that secret is Flow B CI / gate messaging only. Flow B optional companion (not a plan task): when the plan's PR has been pushed and CI has posted its review, the developer MAY invoke the upstream apply-review sub-skill from within the same execute session to walk through CI findings per-finding (apply / defer / skip) with explicit consent — read-only by default, edits require per-finding yes, never commits or pushes. This is surfaced as an available option during execute; the addon MUST NOT insert an apply-review task file into any plan (would violate the mandatory-final-task-order rule).

• Skills & Agents Discovery (MANDATORY, second-to-last, task N-1, {N-1}.task_skills_agents_discovery.md): reviews completed tasks for new patterns, checks the catalog, creates/updates skills/agents if warranted, evaluates the skills-generator system. If no skills-generator exists in the target repo, write a simplified evaluate-and-document-only version.

• Executive Report (MANDATORY, last, task N, {N}.task_executive_report.md): generates analysis_results/EXECUTIVE_REPORT.md with Executive Summary, Product Impact, Technical Details, QA Verification Guide, FAQs, Next Steps.

Add to the README a note: "Every plan includes Security Review (third-to-last), Skills & Agents Discovery (second-to-last), and Executive Report (last). These are auto-generated by /dwp-create."

Quality gates: atomic, ordered tasks; both mandatory files present; analysis_results/ exists; PROGRESS.md exists; numbering correct (user tasks → security review → skills discovery → executive report); behavior-changing tasks carry test coverage in their Acceptance Criteria and tests + lint/type-check in their Validation (../guide/GUIDE.md §5.3).

Accelerate generation with team agents (Claude Code only, automatic): for 5+ user-defined task files, the lead creates README/PROMPTS/PROGRESS/ analysis_results + the mandatory final task files, then spawns teammates (1 per 2–3 task files) to write user task files in parallel (no file overlap), verifies all files, and cleans up. Fallback: generate sequentially.

Orchestrator additions (orchestrator plans):

• README Child DWP Plans table (# | Repository | Child Plan | Status | Depends On), an Execution Mode subsection, and Dependency Rules.
• ORCHESTRATOR_MANIFEST.md at .dwp/plans/PLAN_{name}/ORCHESTRATOR_MANIFEST.md (template in ../guide/GUIDE.md §13.8): Shared Context, Child DWP Registry, Dependency Graph, Output Contracts, Execution State.
• Task files: direct design tasks first (if hybrid); then create_child_dwp tasks in dependency order (template ../examples/ORCHESTRATOR_TASK_TEMPLATE_create_child_dwp.md); optional integration_checkpoint tasks (../examples/ORCHESTRATOR_TASK_TEMPLATE_integration_checkpoint.md); execute_child_dwp tasks for Sequential-with-Output-Handoff mode (../examples/ORCHESTRATOR_TASK_TEMPLATE_execute_child_dwp.md); then the two mandatory final tasks last. Each create_child_dwp task instructs the agent to navigate to the target repo, read its AGENTS.md, and create a child DWP at repositories/{repo}/.dwp/plans/PLAN_{child}/ using that repo's conventions.

Team-agents metadata (when Step 2.10 detected parallelizable tasks): add a "Team Agents Configuration (Claude Code Only)" section to the README (Parallel Task Groups + Teammate Roles tables) and a "Team Agents Metadata (Claude Code Only)" section to each parallel task file (Parallel Group / Teammate Role / Can Run With / Blocks / Files Owned). Use ../examples/TEAM_AGENTS_TASK_TEMPLATE.md. Rules: never put required info inside team-agents sections; every task must work sequentially; mandatory final tasks are always sequential; file ownership between parallel tasks must not overlap.

Step 5 — Completion & Execute Option

For a full plan, report success and the location .dwp/plans/PLAN_{name}/, then offer: (1) execute now → run the Execute sub-skill (../execute/SKILL.md); (2) review the README first, then ask again; (3) done for now → tell them to run /dwp-execute {name} later.

Dailybot kickoff (only when the Dailybot addon is wired — best-effort, non-blocking): after the plan is materialized and approved, send a regular (non-milestone) kickoff report via the dailybot report sub-skill — "Starting: <what is being built and why it matters>" — per the addon's lifecycle event model (../addons/dailybot/SPEC.md §5.1). One kickoff per plan; skip silently if Dailybot is absent, unauthenticated, or .dailybot/disabled exists. Never block on this.

For refined-draft-only mode, report the single refined draft saved at .dwp/drafts/PLAN_{name}_draft_refined.md and the next step: run /dwp-create from PLAN_{name}_draft_refined.md to build the final plan.

Step 6 — Execute Plan (Optional)

If the user chose to execute, hand off to the Execute sub-skill (../execute/SKILL.md): read the plan README, check git status, start from the first [ ] task, execute sequentially, validate, commit per task, report.

Error Handling

• Plan name already exists: offer different name / overwrite / cancel.
• Name auto-converted: show an informational notice (not an error).
• Refined draft not found (from-refined-draft): list available refined drafts in .dwp/drafts/ and ask the user to choose.
• Insufficient tasks (<2): ask the user to break the work down.

Important Notes

• Git ignore: everything under .dwp/ is git-ignored.
• Single artifact: the only draft file is PLAN_{name}_draft_refined.md in .dwp/drafts/ — there is no separate raw draft.
• Reference: follow ../guide/GUIDE.md; orchestrator → §13; team agents → §14.
• Archetypes: orchestrator support assumes the orchestrator-hub archetype (sub-repos under repositories/); an individual repo creates standard plans. See ../shared/adaptation.md.