spears

name	spEARS
description	spEARS (Simple Project with EARS) is a requirements-first specification methodology for AI agents and humans. It pairs timeless EARS requirements (requirements.md) with point-in-time design decision records (ADRs), an on-demand precise behavioral layer (Allium), and a status doc (executive.md), giving a clear shared understanding of how a system should work with grep-able traceability from requirement to test to code. Use this skill whenever the user mentions spEARS or EARS specs; works with a specs/ directory or files named requirements.md, executive.md, or adrs/; asks to write, plan, or validate requirements; wants to capture an architecture or design decision; or asks how requirements trace to tests and code — even if they don't say "spEARS" explicitly.
auto_trigger	[{"file_patterns":["**/requirements.md","**/executive.md","**/adrs/*.md"]},{"keywords":["spEARS","spears spec","EARS requirement","requirements.md","ADR","design decision record"]}]

spEARS: Simple Project with EARS

spEARS is a lightweight, requirements-first methodology for building software with AI agents. Its north star: be explicit about design, so that humans and agents share a clear, accurate understanding of how the system should work — and keep that understanding honest as the code evolves.

A feature is described by four kinds of artifact, each matched to a different relationship with time: timeless requirements (what, and why a user cares), point-in-time ADRs (why a decision was made), an on-demand precise Allium spec (how it behaves now), and an executive status doc (where things stand). Matching each kind of knowledge to a medium that fits it is the whole idea — and it is why spEARS has no always-current “design document,” the one artifact that reliably goes stale. The reasoning is the heart of the method; read references/design-philosophy.md once before working with spEARS in earnest.

Core principles

1. Requirements first. Define what needs to exist, and why a user cares, before writing tests or code.
2. Match the medium to the job. Timeless requirements in prose; precise current behavior in a checkable DSL (Allium); point-in-time decisions in prose ADRs. Never ask a prose document to be a live mirror of running code.
3. Immutable traceability. Requirements get permanent IDs (REQ-XX-###) that thread through specs, tests, and code and never change.
4. YAGNI made detectable. Over-engineering has a precise definition here: code that cannot be traced back to a REQ-ID. That makes it something you can find, not just scold.
5. Specs stay timeless; status stays separate. Every artifact except executive.md describes a standing ideal, as if it had always been that way — no changelogs, no “we recently switched,” no status tables.

The artifact model

spEARS uses four kinds of artifact, each matched to a different relationship with time. Feature specs hold their requirements, status, and (on demand) behavior; ADRs live in one shared, project-level chain, because design decisions routinely cross features.

specs/
├── adrs/                        # ONE shared, cross-cutting ADR chain (sibling of features)
│   ├── _TEMPLATE.md
│   ├── README.md                # index: quick-ref table + task→ADR routing + dependency graph
│   └── NNN_<slug>.md            # WHY — context, options, the call — point-in-time, frozen
└── feature-name/
    ├── requirements.md          # WHAT + user-facing why — EARS, immutable REQ-IDs, timeless
    ├── feature-name.allium      # HOW, EXACTLY — states/transitions/invariants — present tense, on demand
    └── executive.md             # WHERE ARE WE — status — the one "now" document

An ADR names its scope in an Affects: field (REQ-TA-001, or several), never by which folder it sits in — every ADR shares the one specs/adrs/ directory. The specs/ prefix is a recommended default, not a mandate: docs/specs/… or docs/{adrs, specs/…} are equally fine. The one invariant is that ADRs are a sibling of feature specs, never nested inside a feature.

Document	Contains	Never contains
requirements.md	EARS requirements, user-facing rationale, REQ-IDs	Status, implementation detail, decision logs
adrs/*.md	One decision: context, options weighed, the call, consequences	Timeless rule claims, Allium syntax, feature-implementation status
*.allium	Allium declarations of current behavior (normative)	Defer to Allium conventions
executive.md	Status table, brief summaries, verification coverage	Code blocks, decision logs, path-dependent narrative

What is normative (the code must obey it): requirements.md and .allium. If the code disagrees with either, something is wrong — fix the code, or, if the spec is wrong, correct it deliberately (a real change of decision earns an ADR). Do not silently “fix” the spec to match the code. ADRs are authoritative history, not a third contract; you read them to understand intent, not to check compliance.

For the full reasoning on why each artifact has the temporal character it does, see references/design-philosophy.md.

When to reach for Allium

Allium is a heavier, precise tool. It is precision-on-demand, not required on every feature. Reach for a .allium file when the feature has real behavioral complexity:

• State machines — multiple states with non-trivial transitions
• Lifecycle flows — preconditions that must hold (approve, complete, abandon)
• Multi-step operations — ordering matters and partial failure is possible
• Cross-boundary contracts — two specs interact

Skip Allium when requirements.md already says enough: CRUD endpoints, pure data transformations, UI components, tools with no lifecycle. Adding Allium there is cost without payoff. When you do write one, defer to the allium skill for syntax and the elicit / propagate / weed workflow.

spEARS is complete on its own. Allium is an optional companion for the complex minority of features — install it when you want formal behavioral specs and generated tests. Everything else (requirements, ADRs, executive status, open questions, validation) works with spEARS alone. The references note the few places that gain extra power when Allium is present.

The workflow: layered handoff

spEARS owns the requirements → ADR → executive markdown layer and hands off to Allium at the behavioral-spec boundary. The REQ-ID is the suture running through every layer.

discover ──► requirements.md (REQ-IDs, EARS, user-why) + executive.md (skeleton, all ❌)
   │
   ▼  GATE: state-machine-complex? (lifecycle / preconditions / ordering / cross-boundary)
   │
   ├─ NO  ──► implement directly against REQ-IDs, stay in spEARS
   │
   └─ YES ──► allium elicit ─► feature.allium (references the REQ-IDs)
              allium propagate ─► tests (must fail first)
              implement ─► code with // REQ-XX-### comments
              allium weed ─► .allium ↔ code divergence check
   │
   ▼
write an ADR in specs/adrs/ for any significant decision made along the way, and for
every REQ deprecation (Affects: names the REQ-IDs it touches)
   │
   ▼
validate (markdown layer) ──► executive.md status  ❌ → 🔄 → ✅

Routing table

This SKILL.md is the hub. Each task below has a home — a reference file in this skill, or the sibling allium skill. Read the reference when you are doing that task.

Task	Go to	When
Understand why spEARS is shaped this way	references/design-philosophy.md	Before working with spEARS the first time
Discover user need, write requirements	references/discovery.md	New feature, or an existing spec feels hollow
Write EARS requirements correctly	references/ears-guide.md	Authoring or reviewing requirements.md
Author/update requirements & executive	references/authoring.md	Writing the markdown layer
Capture a design decision	references/adr-guide.md	A decision has a real “why”; any REQ deprecation
Specify exact behavior, generate tests	allium skill (elicit, propagate)	The Allium gate fired YES
Validate the markdown layer	references/validation.md	Before merge
Check spec ↔ code divergence	allium skill (weed)	The feature carries a .allium
Establish/verify traceability	references/traceability.md	Linking requirement → test → code
See a full feature worked end-to-end	references/worked-examples.md	You want a concrete model to follow

Traceability in one line

Every requirement gets an immutable REQ-XX-### that you can grep across the whole system: requirements.md defines it, executive.md tracks it, .allium references it, ADRs cite the ones they affect, code and tests carry it in comments. One rg REQ-RL-001 shows you the requirement, its status, its precise behavior, the decisions behind it, and every line of code that implements it. Details and verification commands in references/traceability.md.

A few rules worth stating up front

These are the ones that, when violated, quietly corrupt the system:

• REQ-IDs are immutable. Never renumber, never reuse. Deprecate (don’t delete), and back every deprecation with an ADR.
• Status lives only in executive.md. Not in requirements, not in ADRs, not in Allium.
• Specs are timeless; ADRs are not. Keep dates, “we decided,” and option-weighing inside ADRs and out of everything else. When you resolve a question, state the outcome as a standing fact in the spec and let the ADR hold the deliberation.
• Don’t invent scope. Code with no REQ-ID behind it is the definition of over-engineering here. If a capability is worth building, it is worth a requirement first.

References

• references/design-philosophy.md — why there is no design.md, and the temporal model behind the four artifacts
• references/ears-guide.md — the EARS format, patterns, and good vs. bad requirements (stage 2)
• references/authoring.md — writing requirements.md and executive.md; document-separation rules (stage 2)
• references/adr-guide.md — ADR format, lifecycle, supersession, the shared-chain model, and the index pattern. The format lives in adrs/_TEMPLATE.md; adrs/README.md is the worked index; and adrs/000_no-design-document.md is spEARS’s own founding ADR — the worked exemplar, and proof the method is built with itself.
• references/discovery.md — Socratic discovery of user need (stage 3)
• references/validation.md — validating the markdown layer (stage 3)
• references/traceability.md — grep-based traceability and verification (stage 3)
• references/worked-examples.md — full features worked end-to-end (stage 3)

Sibling skill: allium — the precise behavioral layer. spEARS without Allium is vague about exact behavior; Allium without spEARS is unmoored from user need.