Spec Coder

Spec Coding Workflow

Write specs first, then generate code. 5 phases, each producing artifacts that feed the next. Expert review gates between phases catch issues early.

Session Start

On every new session:

1. Read specs/status.md to identify current phase, active changes, and deferred items.
2. Read phase-relevant spec files — see Quick Reference table below for which artifacts each phase produces/consumes.
3. Confirm with the user which phase/gate to continue from. If mid-phase, summarize progress so far.
4. No specs/status.md? This is a new project — start from Phase 0 (existing codebase) or Phase 1 (greenfield).

Workflow Overview

Phase 0 ──→ Phase 1 ──→ ║Gate 1║ ──→ Phase 2a ──→ Phase 2b ──→ ║Gate 2║
(scan)      (clarify)    (req.)       (design)      (preview)     (design)
                                                                     │
    ┌────────────────────────────────────────────────────────────────┘
    ▼
Phase 2c ──→ Phase 2d ──→ ║Gate 3║ ──→ Phase 3 ──→ ║Gate 4║ ──→ Phase 4 ──→ Phase 5
(tasks)      (specs)       (plan)       (code)       (code)       (verify)    (iterate)
                                                                                  │
                                                                    ┌─────────────┘
                                                                    ▼
                                                              Re-trigger Gate
                                                              2 / 3 / 4 based
                                                              on change scope

Gates auto-approve when no Critical/Major issues. See Expert Review Protocol for details.

Quick Reference

When to Use

Trigger conditions:

• User mentions "spec coding", "spec-first", "spec-driven development"
• User asks to write specs / requirements / design docs before implementation
• User mentions phase keywords: "clarify requirements", "write a spec", "generate from spec"

When NOT to use: Pure bug fixes, one-line config changes, dependency updates, or tasks completable in < 15 minutes without design decisions.

Complexity triage — choose the right track:

Ask the user which track fits, or infer from the scope of their request.

Small track shortcut: Skip Phase 1. Write a single spec_xxx.md (Interface + Business Rules + Test Points only), then go straight to Generate → Verify.

Medium track shortcut: Phase 1 is a brief bullet-point confirmation (no full requirements.md). Phase 2 produces tasks.md + spec_xxx.md only (skip design.md and design-preview/ unless architecture decisions are needed).

Mid-project entry: If the user already has spec files or an existing codebase, read them first. Validate existing artifacts against the expected format, then confirm which phase to start from.

File Organization

Two-layer structure: trunk (current system truth) + changes (incremental work).

specs/
├── index.md                    ← navigation hub: all features & recent changes
├── requirements.md             ← project-level requirements (grows incrementally)
├── design.md                   ← system-level architecture (grows incrementally)
├── design-preview/
├── tasks.md                    ← active tasks only
├── status.md
├── spec_<feature_name>.md      ← feature specs (current truth, one per feature)
└── changes/                    ← all incremental work
    ├── FEAT-NNN-name/          ← new feature (full spec lifecycle)
    │   ├── spec.md
    │   ├── design.md
    │   ├── tasks.md
    │   └── delta.md            ← merge instructions for trunk
    ├── CHG-NNN-name/           ← change/enhancement (delta only)
    │   ├── spec.md             ← ADDED / MODIFIED / REMOVED sections
    │   ├── tasks.md
    │   └── delta.md
    └── archive/                ← completed & merged changes

First-time projects: Start with trunk only (no changes/ needed). The changes layer is introduced when the first post-v1 feature or modification begins.

For full lifecycle management details, see Spec Lifecycle Management.

Phase 0 (Optional): Codebase Scan

For existing codebases, run before Phase 1: read dependency files to detect tech stack, list directory tree, identify existing interfaces/models/conventions, and summarize findings.

Output: Write results to specs/status.md under ## Codebase Context (tech stack, key conventions, existing interfaces, directory structure summary). This persists the scan across sessions. Phase 1 references this for requirements scoping; Phase 2a references it for architecture decisions.

Phase 1: Clarify & Scope

Goal: Turn informal requirements into a confirmed, structured requirements.md.

Constraints: Do NOT write any code. Only ask questions and produce documents.

1. Interactive Clarification — Ask user for raw requirements. Summarize goals (3–5 bullets), list ambiguities with options, suggest unconsidered constraints. Iterate until confirmed.
2. Structured Requirements — Produce specs/requirements.md (see template): Background & Objectives, User Roles & Use Cases, Functional Requirements (FR-001...), Non-Functional Requirements, Out of Scope. If Phase 0 ran, populate the "Existing Architecture" section with requirement-relevant context from status.md § Codebase Context.

→ Gate 1: Requirements Review

Per Expert Review Protocol — Gate 1. Exit: User approves (or auto-approved).

Phase 2: Spec

Goal: Produce spec documents that directly drive code generation.

2a: Technical Design → specs/design.md

Based on requirements.md, produce specs/design.md (see template):

1. Architecture overview (tech stack, layers, services)
2. Core module breakdown with responsibilities and dependency directions
3. Key data models, interfaces & interaction flows
4. Cross-cutting concerns (error handling, auth, observability, deployment) — skip on Small track
5. NFR fulfillment matrix + key architecture decisions (ADR format) — skip on Small/Medium track

For existing codebases: populate "Existing Architecture" section with architecture-relevant context from status.md § Codebase Context (what's new vs. modified).

2b: Design Preview → specs/design-preview/

Self-contained HTML prototype visualizing architecture, module layout, UI screens or API flows, and data models.

Constraints:

• Single HTML file per page/screen (inline CSS + JS, no external dependencies) — must open in any browser without a build step.
• Target: validate information architecture and user flow, not pixel-perfect design. Lo-fi is acceptable.
• Limit to ≤ 5 key screens/pages. Prioritize primary user workflow.
• For non-UI projects: Architecture Diagram Preview only (module relationships + data flow as a single HTML page).

UI projects: All generated UI must follow the UI Design Guidelines to avoid AI-template aesthetics.

→ Gate 2: Design Review (2a + 2b combined)

Per Expert Review Protocol — Gate 2. Exit: User approves (or auto-approved).

2c: Task Breakdown → specs/tasks.md

Task table (see template). Target granularity: 30–90 min per task.

2d: Feature Spec → specs/spec_<feature>.md

Per feature (see template): Feature Description & Use Cases, Interface Definition, Data Model, Business Rules & Edge Cases, Test Points (≥ 5 per feature, complex: 10–20; categories: Normal / Error / Boundary / Combination).

→ Gate 3: Implementation Plan Review (2c + 2d combined)

Per Expert Review Protocol — Gate 3. Exit: User approves (or auto-approved).

Spec Quality Checklist

Before leaving Phase 2: every interface has explicit I/O types; every business rule has ≥ 1 test point; specs describe what not how; each spec is self-contained for code generation; edge cases are explicit, not implied.

Phase 3: Generate

Goal: Produce implementation code + tests strictly based on the spec.

Spec adherence — tiered rules:

Never silently deviate from Strict-tier items. If a Strict item is wrong or incomplete, flag it and ask the user — but do not block on every minor gap.

Execution Strategy

1. Order: Sort tasks by dependency graph (topological order). Infrastructure/setup tasks first.
2. Parallelism: Tasks with no mutual dependencies may be executed in parallel (e.g., via subagents). Tasks sharing the same module should be sequential to avoid merge conflicts.
3. Commit cadence: One commit per task, tagged with task ID: feat(TASK-001): ...
4. Failure handling: If a task fails to generate or test correctly after 2 attempts, mark it Blocked in tasks.md, log the reason, and continue with non-dependent tasks. Return to blocked tasks after unblocked tasks complete.
5. Progress tracking: Update tasks.md Status column (Pending → In Progress → Done / Blocked) as each task progresses.

Steps

1. Scan project structure and tech stack (or use Phase 0 results).
2. For each task in tasks.md (per execution strategy above), take its linked spec_xxx.md as input.
3. Output per task:
   • File list (create vs. modify, with paths)
   • Implementation code
   • Tests derived from spec Test Points (TP-001 → test function)
   • Human-review points — Flexible-tier deviations + all [SPEC-GAP] items
4. Add spec traceability comments at module/class level: // SPEC: spec_auth.md | TASK-002 — enables reverse lookup from code to spec.

For existing codebases: clearly distinguish "new file" vs. "modify existing file" and show diffs for modifications.

→ Gate 4: Code Review

Per Expert Review Protocol — Gate 4. Exit: User approves (or auto-approved). Proceed to Phase 4.

Phase 4: Verify

Goal: Confirm generated code satisfies the spec through execution (static compliance is already covered by Gate 4).

4a: Dynamic Verification

Run the following checks in order. Stop and fix if any step fails.

1. Build — Project compiles / bundles without errors.
2. Lint & Type Check — Run linter and type checker (if available). Zero errors required; warnings acceptable.
3. Unit Tests — Execute test suite (pytest, npm test, go test, etc.). Map each TP-ID to pass/fail.
4. Integration Tests — Run integration / E2E tests if defined in tasks.md.
5. Coverage Check — Report test coverage. Flag spec Test Points that have no corresponding test execution.
6. NFR Verification — If requirements.md defines measurable NFRs (response time, throughput), run benchmarks and compare. Log results even if not blocking.
7. Manual Test Checklist — For tasks marked AI-Auto: No in tasks.md, generate a manual test checklist with steps and expected results for the user to execute.

4b: SPEC-GAP Resolution

For each [SPEC-GAP] from Phase 3: Accept (update spec) / Reject (fix code) / Defer (add to backlog).