Gate 8: Zero-context implementation steps - 2-5 minute atomic subtasks with complete code, exact commands, TDD pattern. Large Track only.
Inherits all available tools
Additional assets for this skill
This skill inherits all available tools. When active, it can use any tool Claude has access to.
name: pre-dev-subtask-creation description: | Gate 8: Zero-context implementation steps - 2-5 minute atomic subtasks with complete code, exact commands, TDD pattern. Large Track only.
trigger: |
skip_when: |
Write comprehensive implementation subtasks assuming the engineer has zero context for our codebase. Each subtask breaks down into 2-5 minute steps following RED-GREEN-REFACTOR. Complete code, exact commands, explicit verification. DRY. YAGNI. TDD. Frequent commits.
Save subtasks to: docs/pre-dev/{feature-name}/subtasks/T-[task-id]/ST-[task-id]-[number]-[description].md
Every subtask must be completable by anyone with zero context about the system.
Requiring context creates bottlenecks, onboarding friction, and integration failures.
Subtasks answer: Exactly what to create/modify, with complete code and verification. Subtasks never answer: Why the system works this way (context is removed).
Each step is one action (2-5 minutes):
Header (required):
| Field | Content |
|---|---|
| Title | # ST-[task-id]-[number]: [Subtask Name] |
| Agent Note | > **For Agents:** REQUIRED SUB-SKILL: Use executing-plans |
| Goal | One sentence describing what this builds |
| Prerequisites | Verification commands with expected output |
| Files | Create: exact/path, Modify: exact/path:lines, Test: tests/path |
Step Structure (TDD Cycle):
| Step | Content |
|---|---|
| Step 1: Write failing test | Complete test file with imports |
| Step 2: Run test to verify fail | Command + expected failure output |
| Step 3: Write minimal implementation | Complete implementation file |
| Step 4: Run test to verify pass | Command + expected success output |
| Step 5: Update exports (if needed) | Exact modification to index files |
| Step 6: Verify type checking | Command + expected output |
| Step 7: Commit | Exact git commands with message |
| Rollback | Exact commands to undo if issues |
Exact file paths (absolute or from root), complete file contents (if creating), complete code snippets (if modifying), all imports and dependencies, step-by-step TDD cycle (numbered), verification commands (copy-pasteable), expected output (exact), rollback procedures (exact commands), prerequisites (what must exist first)
Placeholders: "...", "TODO", "implement here"; vague instructions: "update the service", "add validation"; assumptions: "assuming setup is done"; context requirements: "you need to understand X first"; incomplete code: "add the rest yourself"; missing imports: "import necessary packages"; undefined success: "make sure it works"; no verification: "test it manually"
| Excuse | Reality |
|---|---|
| "The developer will figure out imports" | Imports are context. Provide them explicitly. |
| "TODO comments are fine for simple parts" | TODOs require decisions. Make them now. |
| "They'll know which service to update" | They won't. Specify the exact file path. |
| "The verification steps are obvious" | Obvious ≠ documented. Write exact commands. |
| "Rollback isn't needed for simple changes" | Simple changes fail too. Always provide rollback. |
| "This needs system understanding" | Then you haven't removed context. Simplify more. |
| "I'll provide the template, they fill it" | Templates are incomplete. Provide full code. |
| "The subtask description explains it" | Descriptions need interpretation. Give exact steps. |
| "They can look at similar code for reference" | That's context. Make subtask self-contained. |
| "This is too detailed, we're not that formal" | Detailed = parallelizable = faster. Be detailed. |
| "Steps are too small, feels like hand-holding" | Small steps = verifiable progress. Stay small. |
If you catch yourself writing any of these in a subtask, STOP and rewrite:
..., // TODO, // implement X hereWhen you catch yourself: Expand the subtask until it's completely self-contained.
| Category | Requirements |
|---|---|
| Atomicity | Each step 2-5 minutes; no system architecture understanding required; assignable to anyone |
| Completeness | All code provided in full; all file paths explicit; all imports listed; all prerequisites documented; TDD cycle followed |
| Verifiability | Test commands copy-pasteable; expected output exact; commands run from project root |
| Reversibility | Rollback commands provided; rollback doesn't require system knowledge |
Gate Result: ✅ PASS → Ready for implementation | ⚠️ CONDITIONAL (add details) | ❌ FAIL (decompose further)
| Factor | Points | Criteria |
|---|---|---|
| Step Atomicity | 0-30 | All 2-5 minutes: 30, Most sized right: 20, Too large/vague: 10 |
| Code Completeness | 0-30 | Zero placeholders: 30, Mostly complete: 15, Significant TODOs: 5 |
| Context Independence | 0-25 | Anyone can execute: 25, Minor context: 15, Significant knowledge: 5 |
| TDD Coverage | 0-15 | All RED-GREEN-REFACTOR: 15, Most have tests: 10, Limited: 5 |
Action: 80+ autonomous | 50-79 present options | <50 ask about structure
After creating subtasks, offer execution choice:
"Subtasks complete. Two execution options:
subagent-driven-developmentexecuting-plansWhich approach?"
If you wrote a subtask with "TODO" or "..." or "add necessary imports", delete it and rewrite with complete code.
Subtasks are not instructions. Subtasks are complete, copy-pasteable implementations following TDD.
Every subtask must be completable by someone who:
If they can't complete it with zero questions while following RED-GREEN-REFACTOR, it's not atomic enough.
Remember: DRY. YAGNI. TDD. Frequent commits.