t-plan

T-Plan Skill (Task-Based Orchestration)

MANDATORY: STOP AND READ THIS FIRST

You are an ORCHESTRATOR, not an implementer.

Before doing ANYTHING else, you MUST:

1. Create the session directory with mkdir -p .t-plan/<session-id>
2. Write intent.md capturing the user's request
3. Create the master task with TaskCreate

You MUST NOT:

• Use Read, Grep, Glob, or WebSearch yourself (that's the EXPLORE subagent's job)
• Research documentation yourself (that's the VALIDATE subagent's job)
• Skip directly to analysis or recommendations

If you find yourself exploring the codebase before writing intent.md, STOP. You have violated the workflow.

The user invoked /t-plan because they want the FULL orchestrated workflow with artifacts, not ad-hoc help. Follow the steps below EXACTLY.

---

Transform conversations into rock-solid implementation plans using Claude Code's native Task tools for coordination.

Architecture Overview

No hooks required - Task tools handle all coordination natively.

Orchestrator responsibilities:
- Session initialization (mkdir, .gitignore, current.txt pointer)
- Task lifecycle (TaskCreate, TaskUpdate)
- Pre-truncation before subagent dispatch
- Output verification after subagent return
- Retry logic (max 2 attempts)

Subagent responsibilities:
- Read context files (intent.md, explore.md, etc.)
- Write output files directly (explore.md, scout.md, validation-vNNN.json)
- Focus on their assigned task

Reference material (load when needed):

• Retry logic & parallel execution: See references/subagent-patterns.md
• Planning principles & anti-patterns: See references/planning-principles.md
• Plan output template: See references/plan-template.md

---

Workflow

INTENT -> EXPLORE -> [gate] -> SCOUT -> DRAFT -> VALIDATE -> PLAN
   |          |                   |        |          |         |
 orch.     subagent            subagent  orch.    subagent   orch.+user

Step	Actor	Output
INTENT	Orchestrator	Clear intent (gate: can direct EXPLORE?)
EXPLORE	Subagent	Codebase insights (gate: trivial -> skip?)
SCOUT	Subagent (optional)	Alternatives only (no docs)
DRAFT	Orchestrator	Initial approach (reads files, synthesizes)
VALIDATE	Subagent (required checkpoint)	Doc validation + snippets
PLAN	Orchestrator + User	Final plan, iterate until approved

Principles:

• Subagents gather information
• Orchestrator synthesizes and decides
• Two gates: clarity (INTENT), complexity (EXPLORE)
• Single user checkpoint (PLAN)

---

Session Directory Structure

.t-plan/
+-- current.txt                    # Text pointer: "auth-oauth-20260124-150230"
+-- .gitignore                     # Ignore all session dirs
+-- auth-oauth-20260124-150230/    # Session directory
    +-- intent.md                  # Orchestrator writes (Step 1)
    +-- explore.md                 # EXPLORE subagent writes
    +-- scout.md                   # SCOUT subagent writes (optional)
    +-- draft-v001.md              # Orchestrator writes (Step 4)
    +-- validation-v001.json       # VALIDATE subagent writes
    +-- plan.md                    # Final output (Step 6)

Session ID format: <slug>-<YYYYMMDD-HHMMSS>

---

Step 1: INTENT [ORCHESTRATOR]

Goal: Capture user intent and initialize session.

Initialize Session

SESSION_ID="<slug>-<YYYYMMDD-HHMMSS>"
Bash: mkdir -p .t-plan/${SESSION_ID}
Write: .t-plan/.gitignore -> "*\n!.gitignore\n"
Write: .t-plan/current.txt -> "${SESSION_ID}"
TaskCreate(subject: "T-Plan: [goal]", metadata: {"session_id": "${SESSION_ID}"})
TaskGet + TaskUpdate(status: "in_progress")

Capture Intent

1. Capture the user's request in their terms
2. Clarify ambiguities that would prevent focused exploration
3. Do NOT assume technologies or solutions
4. Write intent to .t-plan/${SESSION_ID}/intent.md

Clarity Gate

"Can I write a focused prompt for the EXPLORE subagent?"

• NO -> Clarify using AskUserQuestion with options: "Narrow scope", "Define outcome", "Add constraints"
• YES -> Proceed to EXPLORE

---

Step 2: EXPLORE [SUBAGENT]

Goal: Understand the codebase relevant to the user's intent.

Dispatch Pattern

# Pre-truncate output
Write: .t-plan/${SESSION_ID}/explore.md -> ""

# Create tracking task
TaskCreate(subject: "EXPLORE: [area]", metadata: {"output_file": "explore.md"})

# Dispatch
Task(subagent_type: "Explore", allowed_tools: ["Read", "Grep", "Glob", "Write"], prompt: """
  **Context**: Read .t-plan/${SESSION_ID}/intent.md

  **Task**: Explore codebase to understand:
  1. Tech stack, frameworks, patterns
  2. Existing code related to [area]
  3. Project structure and conventions
  4. Relevant installed dependencies

  **Output**: Write to .t-plan/${SESSION_ID}/explore.md
  - Start with summary referencing intent.md (proof-of-read)
  - Include key files with line references
  - Describe architecture connections
""")

# Verify: exists + non-empty + references intent.md
# Retry up to 2 attempts (see references/subagent-patterns.md)

Complexity Gate

"Is this task trivially simple?" (single file, pattern exists, no deps, no architecture)

If trivially simple: Ask user: "Continue T-Plan" / "Normal plan mode" / "Skip to DRAFT"

If not trivially simple: Continue to SCOUT or DRAFT.

---

Step 3: SCOUT [SUBAGENT] (Optional)

Goal: Find alternatives that might be simpler than the obvious approach.

Dispatch Pattern

# Pre-truncate + create task
Write: .t-plan/${SESSION_ID}/scout.md -> ""
TaskCreate(subject: "SCOUT: alternatives")

# Dispatch
Task(subagent_type: "general-purpose", allowed_tools: ["Read", "Grep", "Glob", "Write"], prompt: """
  **Context**: Read intent.md and explore.md

  **Task**: Search for alternatives MEANINGFULLY simpler:
  - Fewer dependencies
  - Less code to maintain
  - Better fits existing patterns

  DO NOT query documentation (that's VALIDATE's job).

  **Output**: Write to .t-plan/${SESSION_ID}/scout.md
  Often correct output is: "No simpler alternatives found."
""")

# Verify: exists + non-empty + references explore.md

---

Step 4: DRAFT [ORCHESTRATOR]

Goal: Synthesize all context and draft an initial approach.

What to Do

1. Read intent.md, explore.md, scout.md (if exists)
2. Read key implementation files identified by EXPLORE
3. Write draft to .t-plan/${SESSION_ID}/draft-v001.md

## Approach

**Goal**: [1-2 sentences]

**Key decisions**:
- [Decision 1]: [Choice] because [rationale]. Rejected: [alternatives]

**Files to modify/create**:
- `path/to/file.ts` - [what changes]

**Approach outline**:
1. [High-level step 1]
2. [High-level step 2]

Update master task: TaskUpdate(metadata: {"draft_version": 1})

---

Step 5: VALIDATE [SUBAGENT] (Required Checkpoint)

Goal: Check the draft approach against official documentation.

Note: VALIDATE is required before PLAN. If skipping (internal refactor, no external APIs), record explicit rationale in the plan.

Dispatch Pattern

# Pre-truncate
Write: .t-plan/${SESSION_ID}/validation-v001.json -> ""
TaskCreate(subject: "VALIDATE: draft v1")

# Dispatch
Task(subagent_type: "general-purpose",
     allowed_tools: ["Read", "Grep", "Glob", "Write", "WebSearch", "WebFetch"],
     prompt: """
  **Context**: Read intent.md, explore.md, scout.md (if exists), draft-v001.md

  **Task**: Validate draft against official documentation.
  - Using RECOMMENDED patterns?
  - Any DEPRECATED APIs or anti-patterns?
  - What are the GOTCHAS?
  - Provide WORKING setup snippets

  **Output**: Write to .t-plan/${SESSION_ID}/validation-v001.json
  {
    "draft_version": 1,
    "status": "VALID" | "NEEDS_CHANGES",
    "confirmations": [...],
    "corrections": [...],
    "snippets": [...],
    "gotchas": [...],
    "doc_links": [...]
  }
""")

# Verify: valid JSON + draft_version matches + status exists

Validation Loop

If status: "NEEDS_CHANGES":

1. Revise draft -> draft-v002.md
2. Update metadata: {"draft_version": 2}
3. Re-run VALIDATE with v002
4. Repeat until status: "VALID"

---

Step 6: PLAN [ORCHESTRATOR + USER]

Goal: Present final plan for user approval.

What to Do

1. Incorporate VALIDATE feedback into final approach
2. Enter plan mode: EnterPlanMode()
3. Write .t-plan/${SESSION_ID}/plan.md following references/plan-template.md
4. Present to user, iterate until approved

EnterPlanMode()
Read all context files
Write plan following template
TaskUpdate(status: "completed")

On Approval

Present next steps via AskUserQuestion:

• Execute now: Exit plan mode, begin implementation
• Spawn agent: Create Task with plan.md as context
• Save for later: Report plan location, session complete

---

Quick Reference

Step	Actor	Task Subject	Output
INTENT	Orchestrator	"T-Plan: [goal]"	intent.md
EXPLORE	Subagent	"EXPLORE: [area]"	explore.md
SCOUT	Subagent	"SCOUT: alternatives"	scout.md
DRAFT	Orchestrator	(metadata update)	draft-vNNN.md
VALIDATE	Subagent	"VALIDATE: v{N}"	validation-vNNN.json
PLAN	Orch + User	(task complete)	plan.md

For retry logic and parallel execution patterns, see references/subagent-patterns.md. For planning quality principles and anti-patterns, see references/planning-principles.md.

---

Resume from Artifacts

If session is interrupted and task list is unavailable:

# Find latest session
Read: .t-plan/current.txt -> auth-oauth-20260124-150230

# Check what artifacts exist
Glob: .t-plan/auth-oauth-20260124-150230/*

# Resume from last completed phase:
# - Has intent.md only -> resume at EXPLORE
# - Has explore.md -> resume at complexity gate
# - Has draft-vNNN.md -> resume at VALIDATE
# - Has validation-vNNN.json with VALID -> resume at PLAN