Produces and maintains a holistic system specification (SPEC.md) at a project root. The spec is the single source of truth for what the system must conform to. Use when starting a new project's spec, folding new ADRs or requirements into an existing spec, or reconciling contradicting decisions. Triggers: 'create SPEC.md', 'write a spec', 'system specification', 'update the spec', 'fold ADR into spec', 'specify the system'. Do NOT use for PRDs (run product-definition) or ADRs (run adr-writing); the spec is downstream of both.
Produces and maintains a holistic system specification (SPEC.md) at a project root. The spec is the single source of truth for what the system must conform to. Use when starting a new project's spec, folding new ADRs or requirements into an existing spec, or reconciling contradicting decisions. Triggers: 'create SPEC.md', 'write a spec', 'system specification', 'update the spec', 'fold ADR into spec', 'specify the system'. Do NOT use for PRDs (run product-definition) or ADRs (run adr-writing); the spec is downstream of both.
Produces and maintains a holistic system specification (SPEC.md) at a project root. The spec is the single source of truth for what the system must conform to. Use when starting a new project's spec, folding new ADRs or requirements into an existing spec, or reconciling contradicting decisions. Triggers: 'create SPEC.md', 'write a spec', 'system specification', 'update the spec', 'fold ADR into spec', 'specify the system'. Do NOT use for PRDs (run product-definition) or ADRs (run adr-writing); the spec is downstream of both.
Spec Writing
Produces a single, authoritative SPEC.md describing what the system must conform to right now.
The spec is a snapshot of present truth.
It is not a changelog, not a PRD, not a design doc.
History lives in git.
Decision rationale lives in ADRs.
The spec answers only "what must hold."
Predecessors and Inputs
This skill consumes decided inputs.
It does not interview the user.
Accepted input sources:
Brainstorm transcripts (output of the brainstorm skill)
PRDs (output of the product-definition skill)
ADRs (output of the adr-writing skill)
Loose requirement notes or design decisions
If the input is incomplete or contradictory,
flag the gap and continue with what is decided.
Do not start an inline interview.
Workflow
You MUST follow this sequence.
Skipping a step is a defect, not a shortcut.
Resolve output location — default SPEC.md at repo root.
Accept an override path if the user supplies one.
Read existing spec in full (update flow only) — before composing any changes.
Ingest inputs — read every supplied input source.
Select sections — start from assets/spec-template.md as the canonical skeleton and conditional-section menu.
Pick from the menu based on input shape.
Do not pad with sections the inputs do not warrant.
Strip the template's guidance comments before writing.
Apply voice rules — see references/voice-guide.md.
RFC 2119 normativity, status line, "Implementation-defined" definition,
external protocols referenced not redefined.
Preserve structural elements (update flow only) —
see references/structural-elements.md.
Numbered invariants, named error categories, pseudocode blocks, inline defaults
MUST survive edits unless the user has explicitly directed their change.
Detect and resolve conflicts (update flow only) — see references/update-flow.md.
Any new input contradicting an existing requirement MUST be surfaced
with resolution options. Wait for the user's decision on each contradiction.
Never auto-resolve.
Write — apply the changes to the file directly.
The file is tracked in source control; the user iterates via normal edits
and VCS, not via an in-chat preview of the full spec.
Report — list the sections added or modified in one or two sentences.
Surface any open questions or gaps in the inputs.
Flow-Specific Rules
Bootstrap (no existing spec)
Output a complete spec from scratch.
Section selection is driven entirely by the inputs.
Write the file directly. The user iterates by editing the file or by re-invoking the skill with additional inputs.
Update (existing spec present)
Read the existing spec in full first.
Identify which sections the new inputs touch.
Patch in place; do not regenerate untouched sections.
Status line update: bump only on a substantive change.
Cosmetic edits do not bump status.
Every contradiction with an existing requirement is a conflict (see update-flow.md)
and pauses for user resolution before the write.
Severity-specific resolution options live in update-flow.md.
Hard Gates
The following MUST hold for every invocation:
No silent supersession. A new input contradicting an existing requirement MUST be surfaced as a conflict before the write. The skill MUST present the user with resolution options. The skill MUST NOT pick a side.
No structural element rewrite as a side effect. Numbered invariants, named error categories, pseudocode blocks, and inline defaults change only when the user has explicitly directed the change.
No invented sections. Only emit section types listed in assets/spec-template.md. If a needed section type is missing, propose adding it to the template before using it.
No in-chat preview of the full spec before writing. The file is tracked; the chat is for conflict resolution and post-write reporting, not for previewing content the user can read on disk.
Output Contract
The produced spec MUST satisfy all of:
Begins with a title and a status line (e.g. Status: Draft v1).
Defines "Implementation-defined" once at the top, as a term used to mark intentionally open policy choices.
Uses RFC 2119 keywords (MUST, MUST NOT, SHOULD, SHOULD NOT, MAY, REQUIRED, OPTIONAL, RECOMMENDED) for every normative statement.
Has explicit Goals and Non-Goals sections.
References external protocols and standards by name; does not restate them.
Contains no changelog, revision history, or version table.
Describes the system at the contract level, not the implementation level.
Internal packages, modules, language-specific types, DDL, and infrastructure
technologies MUST NOT appear in normative prose unless they are themselves
the public contract. See references/voice-guide.md
under "Level of Description".
The full set of voice and structural rules lives in the references.
Validation Checklist
Before writing, verify:
Status line present.
"Implementation-defined" defined once at top.
RFC 2119 keywords used for every normative statement; not as English synonyms.
Goals and Non-Goals both present.
No internal package, module, file, or directory path used as a canonical identifier in normative prose.
No language-specific types in field signatures; domain model uses language-neutral types only.
No DDL, table names, columns, or grants in normative prose unless the database schema is the public contract.
No internal infrastructure technologies named in requirements about internal coupling.
Test matrix and Implementation Checklist describe observable behaviors, not code-path or module deliverables.
Each requirement states an externally observable invariant, not the mechanism (cache ordering, upsert flag, index/constraint shape, transaction boundary) the implementation uses to satisfy it.
Non-functional solutions (caching, indexing, batching, queueing) appear only where the externally observable property they serve (latency, atomicity scope, ordering guarantee, throughput target) is also specified.
When the system has more than one component, a System Overview / Components section precedes the Core Domain Model and names components by role.
Every section emitted is listed in the spec template.
No section omitted because of laziness; only because its when-to-include criteria do not apply.
(Update) Every structural element from the previous spec is present in the new spec, identically identified, unless the user directed a change.
(Update) Every conflict between new input and an existing requirement has been surfaced and resolved before the write.
No changelog, no version table, no in-document history.
Anti-Patterns
Producing a README- or design-doc-flavored document with "must" used as English. Use RFC 2119 keywords as graded normativity.
Padding the spec with section types the inputs do not warrant (a stateless library has no state machine).
Regenerating the whole spec on update; only patch sections the inputs touch.
Silently picking a side when a new input contradicts an existing requirement.
Bumping the status line on cosmetic edits.
Embedding a changelog or version history in the spec body.
Restating an external protocol the spec only needs to reference.
Smuggling implementation choices into normative prose: internal package paths, language-specific types, DDL for internal storage, named internal infrastructure. The spec describes the contract; implementation belongs in code and design docs.
Specifying mechanism (cache ordering, upsert with flag, index/constraint shape, transaction boundary) instead of the externally observable invariant the mechanism preserves. Apply the "replace with somehow" litmus test; if the requirement still has meaning, the mechanism was incidental.
When Not to Use
The user has a single architectural decision to record — use adr-writing instead.
The user is describing one product feature — use product-definition instead.
The user wants to verify code conformance to an existing spec — that is a future skill, not this one.
The user has not decided yet — run brainstorm first.
Integration with Other Skills
brainstorm, product-definition, adr-writing produce inputs this skill consumes.
writing-clearly-and-concisely applies to spec prose.
commit handles the commit after the user has reviewed the written file.
Common Issues
Input has no normative structure
Cause: The user passed a brainstorm transcript or a bullet list.
Solution: That is expected. This skill imposes the structure; inputs do not need to.
User asks for a multi-file spec
Cause: A section has grown large enough to feel like its own document.
Solution: Multi-file splitting is out of scope for v1. Keep the spec single-file. Note the size as an open question for a future iteration.
Conflict with no clean resolution
Cause: A new input contradicts an existing requirement in a way that cannot be both-and-ed.
Solution: Surface the conflict with at least three options: keep existing, replace with new, keep both with user-supplied reconciliation. Do not pick.
