Feature planning and implementation with 4 adaptive phases (Specify, Design, Tasks, Execute). Auto-sizes depth by complexity. Writes testable requirements in EARS notation, atomic tasks, atomic Conventional Commits, and requirement traceability. Ships deterministic Python validation scripts so structural gates are enforced by code, not memory. Features an independent Verifier (author != verifier, evidence-or-zero), a discrimination sensor, a decision log (STATE.md), a test-coverage matrix, and a self-improving lessons layer. Stack-agnostic and tool-agnostic. Use when (1) planning features, (2) implementing with verification and atomic commits, (3) validating an implementation against a spec. Triggers on "specify feature", "discuss feature", "design", "tasks", "implement", "validate", "verify work", "UAT", "record decision", "pause work", "resume work". Do NOT use for pure architecture decomposition analysis or standalone technical design documents.
Instrucciones de origen · Vista previa de solo lectura
name
tlc-spec-driven
description
Feature planning and implementation with 4 adaptive phases (Specify, Design, Tasks, Execute). Auto-sizes depth by complexity. Writes testable requirements in EARS notation, atomic tasks, atomic Conventional Commits, and requirement traceability. Ships deterministic Python validation scripts so structural gates are enforced by code, not memory. Features an independent Verifier (author != verifier, evidence-or-zero), a discrimination sensor, a decision log (STATE.md), a test-coverage matrix, and a self-improving lessons layer. Stack-agnostic and tool-agnostic. Use when (1) planning features, (2) implementing with verification and atomic commits, (3) validating an implementation against a spec. Triggers on "specify feature", "discuss feature", "design", "tasks", "implement", "validate", "verify work", "UAT", "record decision", "pause work", "resume work". Do NOT use for pure architecture decomposition analysis or standalone technical design documents.
Loading this skill's files. Reference files live under in this skill's own directory (where this resides). Resolve them relative to the skill directory - never the workspace root - and load them through the active skill by name; never assume a fixed install path. When a step tells you to read a reference, before acting - never act on a partial/truncated read.
references/
SKILL.md
read it completely (to EOF)
Running this skill's scripts. Every scripts/*.py shipped with this skill lives under that same skill directory. Resolve the skill directory first, then invoke python3 <skill-dir>/scripts/<name>.py .... Never run python3 scripts/... from the consuming project root - that looks for a project-local scripts/ tree that is not this skill. Project data under .specs/ is still read/written relative to the project root (pass --root when the cwd is elsewhere). Below, <skill-dir> means the directory that contains this SKILL.md.
Execution contract - every task, non-negotiable (holds even if you do not open the reference files):
Tests derive from the spec's acceptance criteria and assert spec-defined outcomes - they never mirror the implementation.
The gate must pass (tests pass) before a task is done - the test runner decides, not self-assessment.
One atomic commit per task. Mark the task complete in tasks.md (and update spec traceability when used) before that commit, and include those updates in the same commit. Never batch tasks; never weaken, skip, or delete tests to make them pass.
After the LAST task, a fresh Verifier always runs automatically (author ≠ verifier) - spec-anchored outcome check + discrimination sensor. It is never optional and never prompted. See Sub-Agent Delegation.
Blast radius: approving a spec or tasks authorizes local implementation and local commits only. git push, force-push, deploy, production DB changes, and other remote / externally visible / destructive operations require an explicit go-ahead for that action.
Deterministic gates run before human review - not from memory. The structural gates for the spec and tasks are enforced by scripts in this skill's scripts/ directory, so they cannot silently drift when the model forgets a step:
Before confirming a spec: python3 <skill-dir>/scripts/validate_spec.py <spec-path-or-feature> (closure gate: EARS-shaped ACs, filled assumptions, well-formed requirement IDs, required sections).
Before presenting tasks for approval: python3 <skill-dir>/scripts/validate_tasks.py <tasks-path-or-feature> (granularity smell, diagram-vs-Depends on parity within a phase, no forward-phase dependency, every task carries Tests + Gate).
On each commit: python3 <skill-dir>/scripts/check_commit.py --message "<msg>" (Conventional Commits). Optionally wire it as a git commit-msg guard (git only, no agent dependency) - see implement.md.
Before declaring a feature done: python3 <skill-dir>/scripts/validate_state.py <feature> (completion gate: the Verifier's validation.md exists, its verdict is filled to PASS, and it cites file:line evidence - a missing, FAIL, placeholder, or evidence-free report fails). The closing step of Execute runs this automatically, the same way the lessons layer runs at distillation; it is not a manual step.
A non-zero exit means STOP and fix before proceeding. Skip a script only when no code-execution tool is available; then perform the same checks by reading the artifact.
Before Execute: read implement.md completely and run <skill-dir>/scripts/validate_tasks.py; if a formal tasks.md packs into more than one task-budgeted batch (> ~8 tasks), present the sub-agent offer first (see Sub-Agent Delegation).
Auto-Sizing: The Core Principle
The complexity determines the depth, not a fixed pipeline. Before starting any feature, assess its scope and apply only what's needed:
Specify and Execute are always required - you always need to know WHAT and DO it
Design is skipped when the change is straightforward (no architectural decisions, no new patterns)
Tasks is skipped when there are ≤3 obvious steps (they become implicit in Execute)
Discuss is triggered within Specify when the agent detects ambiguous gray areas that need user input, or when the feature has any implicit-requirement dimension present (persistence/state, external calls, auth, payments, concurrency, state transitions)
Interactive UAT is triggered within Execute only for user-facing features with complex behavior
Safety valve: Even when Tasks is skipped, Execute ALWAYS starts by listing atomic steps inline (see implement.md). If that listing reveals >5 steps or complex dependencies, STOP and create a formal tasks.md - the Tasks phase was wrongly skipped.
.specs Structure
.specs/
├── STATE.md # Project memory: Decisions log (AD-NNN) + Handoff snapshot
├── LESSONS.md # Self-improving lessons playbook (rendered by scripts/lessons.py - do not hand-edit)
├── lessons.json # Canonical lessons state (machine-owned)
└── features/ # Feature specifications
└── [feature]/
├── spec.md # Requirements with traceable IDs
├── context.md # User decisions for gray areas (only when discuss is triggered)
├── design.md # Architecture & components (only for Large/Complex)
├── tasks.md # Atomic tasks with verification (only for Large/Complex)
└── validation.md # Verifier report: PASS/FAIL, per-AC evidence, sensor result, diff range
Create artifacts lazily. Write each file only when its phase actually produces content - never scaffold empty context.md, design.md, or tasks.md up front. An empty file signals a phase happened when it did not; absence is the correct state for a skipped phase. The deterministic validators (scripts/validate_spec.py, scripts/validate_tasks.py, scripts/check_commit.py, scripts/validate_state.py) ship inside this skill's own scripts/ directory, alongside lessons.py.
Reconcile Handoff against git (branch, status --porcelain, recent commits) and tasks.md - evidence wins over a stale snapshot. Full procedure: memory.md.
Propose the reconciled next step before writing code.
Context Loading Strategy
On-demand load (only what the current task needs):
.specs/STATE.md - Decisions section (read at Design, re-read on resume); Handoff section (read on resume only)
confirmed lessons - load at Specify and Design via python3 <skill-dir>/scripts/lessons.py list --status confirmed (lessons.md); confirmed only, never candidates
spec.md (when working on a specific feature)
context.md (when designing or implementing from user decisions)
design.md (when implementing from design)
tasks.md (when executing tasks)
Never load simultaneously:
Multiple feature specs
Multiple architecture docs
Target: <40k tokens total context
Reserve: 160k+ tokens for work, reasoning, outputs
Monitoring: Display status when >40k (see context-limits.md)
Sub-Agent Delegation
Trigger: count total tasks. If the feature packs into more than one task-budgeted batch (> ~8 tasks) → offer sub-agents; if it fits a single batch (≤ ~8 tasks) → execute inline.
Offer-then-confirm - never auto-spawn. The user must accept before any sub-agent is dispatched.
One worker per task-budgeted batch (~7 tasks, whole phases): Phases stay the semantic/dependency unit; a batch is the execution unit - one or more consecutive whole phases packed to ~7 tasks. Walk phases in order, accumulate whole phases into the current batch until it reaches the budget, then start the next - never split a phase across workers. ~20 tasks → ~3 workers; scales linearly (40 → ~6). Each worker executes all its tasks in order (implement → gate → atomic commit), then reports a compact summary (tasks done, commit hashes, test counts, deviations). Batches run sequentially - a batch never starts until the previous one reports all tasks complete. Workers never spawn further sub-agents.
Verifier (always-on, never prompted): After the final task is committed, the orchestrator dispatches a fresh Verifier sub-agent automatically - regardless of phase count. Validation never requires a user prompt; it is the closing step of Execute. Author ≠ verifier: the Verifier re-derives coverage independently using evidence-or-zero; it does not inherit the author's mental model. The Verifier: (1) performs a spec-anchored outcome check - confirms each test's asserted value matches the spec-defined expected outcome, flags spec-precision gaps; (2) runs a discrimination sensor - injects behavior-level faults in an isolated scratch (temp worktree or file copies - never git stash), confirms tests kill them, discards the scratch and verifies real-tree porcelain matches the pre-sensor baseline; surviving mutants become fix tasks; (3) writes .specs/features/[feature]/validation.md (PASS/FAIL, per-AC evidence, sensor result, diff range); (4) returns a compact verdict + ranked gap list to the orchestrator in chat. Gaps become fix tasks; the fix→re-verify loop is bounded to 3 iterations before escalating. (5) distills lessons - turns each grounded failure (surviving mutant, spec-precision gap, failed AC, SPEC_DEVIATION) into a reusable project-local lesson via <skill-dir>/scripts/lessons.py; a clean PASS records nothing (see lessons.md).
Model tier per role (only if the harness supports choosing a model per sub-agent). Match the reasoning cost to the work instead of paying top-tier reasoning for boilerplate. A batch worker on a mechanical, low-ambiguity phase (entities, config, wiring, straightforward CRUD) runs on a faster/cheaper tier; a worker on a core-domain or high-ambiguity phase, and the Design phase itself, runs on a high-reasoning tier; the Verifier runs on a mid-to-high tier because it does adversarial reasoning and designs mutations. This is a portable recommendation: if the harness cannot set a per-sub-agent model, ignore it. Full rubric in sub-agents.md.
Standalone fallback: Without sub-agents, run validate.md as an independent fresh-eyes pass after the final commit - including the spec-anchored check and discrimination sensor.
Full mechanics (worker payload, compact summary format, failure handling, context sizing, model tier, Verifier report format): sub-agents.md.
When researching, designing, or making any technical decision, follow this chain in strict order. Never skip steps.
Step 1: Codebase → check existing code, conventions, and patterns already in use
Step 2: Project docs → README, docs/, inline comments, `.specs/STATE.md` (Decisions)
Step 3: Context7 MCP → resolve library ID, then query for current API/patterns
Step 4: Web search → official docs, reputable sources, community patterns
Step 5: Flag as uncertain → "I'm not certain about X - here's my reasoning, but verify"
Rules:
Never skip to Step 5 if Steps 1-4 are available
Step 5 is ALWAYS flagged as uncertain - never presented as fact
NEVER assume or fabricate. If you cannot find an answer, say "I don't know" or "I couldn't find documentation for this". Inventing APIs, patterns, or behaviors causes cascading failures across design → tasks → implementation. Uncertainty is always preferable to fabrication.
Output Behavior
Do the work; do not narrate the machinery. Produce the right artifact for the phase instead of announcing the phase ("I will now run the Specify phase"). The user judges the output, not a play-by-play of the process. This keeps the flow from reading as robotic.
Match effort to the work. Lightweight steps (feature-level checks, validation, mechanical tasks) do not need top-tier reasoning; heavy steps (complex design, ambiguous features) do. If the harness lets you pick a model per sub-agent, apply the tier rubric in sub-agents.md; otherwise proceed and simply invest more care on the heavy steps. Mention this once per session at most, and only if it helps; skip it for an experienced user.
Write generated artifacts in a plain, decided voice. Specs, ADRs, validation reports, commit messages, and chat summaries follow the writing rules in coding-principles.md: lead with the verdict, state decisions definitively, cut filler and mechanical hedging.
Code Analysis
Use available tools with graceful degradation. See code-analysis.md.
