Spec Kit

Overview

A distillation of github/spec-kit — templates and method only, no CLI, no install. Specifications are the primary artifact; code is their expression. You maintain software by evolving the spec, not by patching code and hoping the doc catches up.

The value is not the folder structure. It is that the templates constrain the model: they forbid implementation detail in the spec, force [NEEDS CLARIFICATION] instead of a plausible guess, and gate the plan behind explicit simplicity checks. Fill the templates faithfully and the constraints do the work.

When to Use

• A feature or project needs a durable, reviewable spec that outlives the session
• Requirements are vague, and guessing would cost more than asking
• The work spans multiple sessions, agents, or people and needs one source of truth
• You want traceability: every task back to a requirement, every requirement back to a user story
• Before implementing, to check spec/plan/tasks actually agree

When NOT to use: one-line fixes, typo corrections, exploratory spikes you intend to throw away. For a lighter-weight spec habit without the artifact tree, use spec-driven-development. For task sizing and dependency mechanics, planning-and-task-breakdown is canonical.

Artifacts

Everything for one feature lives in one directory, committed to the repo:

.specify/memory/constitution.md   # project-wide, written once, amended rarely
specs/NNN-feature-name/
├── spec.md                       # WHAT and WHY — no tech stack
├── checklists/requirements.md    # quality gate for spec.md
├── plan.md                       # HOW — stack, structure, constitution gates
├── research.md                   # optional: option comparisons, benchmarks
├── data-model.md                 # optional: entities and relationships
├── contracts/                    # optional: API/event contracts
└── tasks.md                      # executable task list, grouped by user story

NNN is the next free 3-digit number in specs/. Derive the slug from the feature description in action-noun form (add-user-auth, fix-payment-timeout). The spec directory and the git branch name are independent — matching them is a choice, not a rule.

Templates live in templates/ next to this file. Copy, don't paraphrase:

Template	Purpose
templates/spec-template.md	Feature spec: prioritized user stories, FRs, success criteria
templates/plan-template.md	Implementation plan: technical context, gates, structure
templates/tasks-template.md	Task list organized by user story, with [P] parallel markers
templates/constitution-template.md	Project principles that every plan is checked against
templates/checklist-template.md	Generic checklist shell for any quality gate

The Pipeline

CONSTITUTION ─→ SPECIFY ─→ CLARIFY ─→ PLAN ─→ TASKS ─→ ANALYZE ─→ IMPLEMENT
   (once)          │          │         │        │         │
                   └──── quality checklist ──────┴─── read-only audit

Each phase reads the artifacts before it. Do not skip forward; a plan built on an unclarified spec produces tasks that implement the wrong thing efficiently.

0. Constitution (once per project)

Copy templates/constitution-template.md to .specify/memory/constitution.md and fill in the principles this project will not negotiate. Spec Kit ships five example principles — library-first, CLI interface, test-first, integration testing, and one project-defined slot — but the point is that you choose them. What matters is that they are few, concrete, and testable at plan time.

Give it a version and a ratification date. Amendments require a rationale and a compatibility note; principles are meant to be stable across features.

1. Specify

Copy templates/spec-template.md to specs/NNN-slug/spec.md and fill it from the feature description.

Rules that make this phase work:

• WHAT and WHY only. No languages, frameworks, APIs, or code structure. If you catch yourself writing "using React," you are in the plan phase.
• Written for a non-technical stakeholder. They should be able to approve it.
• User stories are prioritized (P1, P2, P3) and independently testable. Each one, built alone, must be a viable slice that delivers value. P1 alone is the MVP.
• Make informed guesses; record them. Reasonable defaults (standard auth, industry retention periods, user-friendly error handling) go in the Assumptions section, not into a question.
• At most 3 [NEEDS CLARIFICATION: specific question] markers. Use them only where no reasonable default exists and the choice moves scope. Priority order: scope > security/privacy > UX > technical detail.
• Success criteria are measurable and technology-agnostic. "Users complete checkout in under 3 minutes" — not "API p95 under 200ms."

Then write specs/NNN-slug/checklists/requirements.md from templates/checklist-template.md with these items, and actually run them against the spec:

## Content Quality
- [ ] No implementation details (languages, frameworks, APIs)
- [ ] Focused on user value and business needs
- [ ] Written for non-technical stakeholders
- [ ] All mandatory sections completed

## Requirement Completeness
- [ ] No [NEEDS CLARIFICATION] markers remain
- [ ] Requirements are testable and unambiguous
- [ ] Success criteria are measurable
- [ ] Success criteria are technology-agnostic
- [ ] All acceptance scenarios are defined
- [ ] Edge cases are identified
- [ ] Scope is clearly bounded
- [ ] Dependencies and assumptions identified

## Feature Readiness
- [ ] All functional requirements have clear acceptance criteria
- [ ] User scenarios cover primary flows
- [ ] Feature meets measurable outcomes defined in Success Criteria
- [ ] No implementation details leak into specification

Fix and re-check, up to three passes. If items still fail, record them in the checklist notes and say so rather than marking them green.

2. Clarify

Before planning, scan the spec for ambiguity across this taxonomy and mark each area Clear / Partial / Missing:

functional scope and out-of-scope · user roles · entities, identity, lifecycle · data volume · critical journeys · error/empty/loading states · accessibility and localization · performance · scalability · reliability · observability · security and privacy · compliance · external services and their failure modes · import/export formats · negative scenarios · rate limiting · concurrency conflicts · technical constraints · rejected alternatives · glossary consistency · acceptance-criteria testability

Ask at most 5 questions, highest impact first, each with 2–4 concrete options and their implications in a table plus a custom option. Present them together, wait for answers, then write the answers back into the spec body — into the requirement or scenario they affect, not into a dangling Q&A appendix.

If the user chooses to skip clarification (a spike), proceed but say plainly that rework risk goes up.

3. Plan

Copy templates/plan-template.md to specs/NNN-slug/plan.md. This is where technology appears for the first time.

Fill Technical Context — language/version, dependencies, storage, testing, target platform, project type, performance goals, constraints, scale. Anything genuinely unknown stays NEEDS CLARIFICATION rather than becoming an invented number.

Then run the Constitution Check gate, deriving the gates from the project's own constitution. Typical shapes:

### Simplicity Gate
- [ ] Using ≤3 projects?
- [ ] No future-proofing?

### Anti-Abstraction Gate
- [ ] Using the framework directly, not wrapping it?
- [ ] Single model representation?

### Integration-First Gate
- [ ] Contracts defined?
- [ ] Contract tests written?

A failed gate is not a blocker — it is a debt you must name. Record it in the Complexity Tracking table with the violation, why it is needed, and why the simpler alternative was rejected. An empty justification means go back and simplify.

Keep the plan high-level and readable. Code samples, detailed algorithms, and long schemas belong in research.md, data-model.md, or contracts/, not inline.

Re-check the gates after the design is filled in. Design drift is the usual way a passing gate quietly becomes a failing one.

4. Tasks

Copy templates/tasks-template.md to specs/NNN-slug/tasks.md. Derive tasks from the plan, the data model, and the contracts — not from imagination.

Structure, which is the whole point of this template:

• Phase 1 Setup — project init, tooling
• Phase 2 Foundational — the blocking prerequisites. No user story starts until this is done.
• Phase 3+ One phase per user story, in priority order. P1 is labeled MVP. Each phase ends with a checkpoint: that story works and is testable on its own.
• Final phase Polish — cross-cutting cleanup, docs, hardening

Task format: [ID] [P?] [Story] Description with exact file path. [P] means it touches different files from its siblings and can run in parallel. [US1] maps it back to a user story for traceability.

Ordering within a story: tests (if requested) → models → services → endpoints → integration. Tests are optional in Spec Kit's template — include them when the spec asks, or when the constitution mandates test-first, in which case they must be written and observed failing before implementation.

5. Analyze

A read-only audit across spec.md, plan.md, and tasks.md. Modify nothing. Produce findings, then offer a remediation plan the user approves before anything is edited.

Build the inventories first — every FR-### and SC-###, every user story, every task and what it maps to — then run the passes:

Pass	Looks for
Duplication	Near-identical requirements that should be consolidated
Ambiguity	"fast", "scalable", "secure", "intuitive" without a number; leftover TODO/???
Underspecification	Requirements with a verb but no measurable object; tasks citing files that appear nowhere in the plan
Constitution alignment	Anything conflicting with a MUST principle
Coverage gaps	Requirements with zero tasks; tasks tracing to no requirement

Severity: constitution conflicts are CRITICAL by definition — resolve them by changing the spec, plan, or tasks, never by reinterpreting the principle. Cap the report at ~50 findings and summarize the overflow.

6. Implement

Work the tasks in order, respecting phase gates and checkpoints. Stop at each user-story checkpoint and verify that story independently before moving to the next priority.

Follow incremental-implementation for the execution loop and test-driven-development when the constitution mandates test-first. Use context-engineering to load only the spec sections and source files each task needs.

When implementation reveals the spec was wrong: update the spec first, then the plan, then the code. Fixing code while leaving the spec stale reinstates exactly the drift this method exists to eliminate.

Why the Templates Work

The templates are prompts that constrain output. Preserve these constraints when you adapt them — dropping them is how the method degrades into ordinary documentation.

1. No premature implementation detail — keeps the spec stable while technology churns
2. Mandatory uncertainty markers — replaces plausible-but-wrong guesses with a question
3. Checklists as unit tests for the spec — forces systematic self-review
4. Gates with justification tables — makes complexity accountable instead of invisible
5. Hierarchical detail — the plan stays readable; depth moves to sub-documents
6. Contracts and tests before source — testability is designed in, not retrofitted
7. No speculative features — every feature traces to a user story with acceptance criteria

See references/philosophy.md for the reasoning behind these.

Red Flags

• Tech stack named in spec.md
• More than 3 [NEEDS CLARIFICATION] markers, or zero in a genuinely vague request
• Success criteria phrased as internal metrics (cache hit rate, TPS) rather than user outcomes
• User stories that cannot be built or tested independently
• tasks.md written from the spec while skipping the plan
• A failed constitution gate with an empty Complexity Tracking row
• analyze editing files instead of reporting
• Code changed to fix a problem the spec still describes incorrectly

Verification

Before implementation:

• constitution.md exists and the plan's gates derive from it
• spec.md passes its requirements checklist with no unresolved markers
• Every user story has a priority and an independent test
• Success criteria are measurable and technology-agnostic
• plan.md records the structure decision and justifies every gate violation
• tasks.md covers every functional requirement, and every task traces to one
• analyze found no CRITICAL findings, or they were resolved in the artifacts
• All artifacts are committed