# Spec > Use when the user explicitly asks for a specification, spec, or contract. Produces a persistent specification with no open decisions in its scope: what, why, observable behavior, constraints, and acceptance criteria. Never produces task breakdowns, estimates, or how-to guidance. - Skill: `harikrishnan94/spec` (Agent Skill) - Install (CLI): `npx skillmds@latest add harikrishnan94/spec` - Raw SKILL.md: https://api.skillmd.com/api/skills/harikrishnan94/spec/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: harikrishnan94 (https://skillmd.com/u/harikrishnan94) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/harikrishnan94/spec --- # Spec generator Input: the user's requested scope or topic, plus any supplied repository, file, product, or conversation context. Output: a **specification** as a persistent contract. ## Artifact - The only artifact is the spec, and it must be persistent. - If the host can edit files and a target path is provided (or a clear convention applies), create or update only that spec file. - Otherwise, output the complete spec as Markdown so the user can persist it. ## What a specification is A spec defines the *what* and the *why* of a change with enough precision that any competent engineer could implement it without making further design decisions. It is a contract on observable behavior, not a sequence of work. ## What a specification is NOT Under no heading, in no section, with no phrasing, do you produce: - Step sequences or ordered procedures - File-by-file or function-by-function change lists - Time, effort, or complexity estimates - Task breakdowns, tickets, or assignee suggestions - Recommendations on *how* to build it - Open questions, TBDs, deferred decisions, or any unresolved choice about the *what* or *why* inside the spec's scope ## Sections (include only those that genuinely apply) - **Problem** — what is true today; why it is unsatisfactory; concrete evidence - **Goals** — observable outcomes that define success - **Non-goals** — explicitly out of scope - **Constraints** — correctness, performance budgets, ABI/API/wire compatibility, build/runtime requirements, security, licensing - **Interfaces & contracts** — signatures, invariants, error semantics, ordering guarantees, observable behavior at boundaries - **Acceptance criteria** — checks that decide whether the spec was met (each independently verifiable) - **Stop conditions** — triggers that require halting and re-opening the spec rather than silently working around it (see below) ## Stop conditions The spec governs the work done against it. Encode explicit triggers under which the implementer halts and re-opens the spec rather than silently works around it. Standard triggers: - A stated constraint cannot be satisfied without violating another constraint or a goal. - A non-goal turns out to be load-bearing for a goal (i.e., a goal cannot be met without doing what was declared out of scope). - An acceptance criterion is not independently verifiable as written. - New evidence contradicts the problem statement or invalidates a constraint. - The interface or behavior contract requires changes not anticipated by this spec. - Scope must expand beyond what the spec authorizes. Specify the response: stop, name the condition, state what changed, propose the minimum revision that would resolve it, and wait for the spec to be updated before continuing. Do not improvise, do not broaden scope silently, do not file a TODO and proceed. ## No open decisions A finalized spec contains no unresolved decisions inside its own scope. Every *what* and *why* the spec governs is settled by the time the spec is persisted. If a decision cannot be settled within the spec, do not finalize the spec and do not paper over the gap. Stop drafting and ask the user how to proceed. Present a small set of concrete, mutually exclusive options — each a fully-formed candidate the spec could adopt — and always include **HALT** as the final option. Each option must specify the decision exactly as the spec would phrase it, the rationale tying it to the established goals and constraints, and the trade-offs it accepts. The **HALT** option means do not finalize the spec until the user supplies more input; state plainly what input would unblock it (missing evidence, a stakeholder ruling, scope clarification, etc.). Resume only once the user has chosen. Do not commit the question as a "TBD," "open question," or "decide later" inside the spec. References to things outside the spec's scope (non-goals, behavior of external systems, future work belonging to a separate spec) are not open decisions, because the spec has explicitly chosen not to govern them. ## Evidence discipline No factual claim about anything outside this conversation may rely on memory or assumption. Before asserting any fact about a library, tool, compiler, API, build flag, codebase symbol, runtime behavior, version, benchmark number, or external standard, retrieve it fresh from an authoritative source: installed tool output (with the exact command and captured result), the codebase at a specific commit (with file:line), or the web (with URL and retrieval date). Pin versions. Cite inline at the claim. If a source cannot be confirmed within the session, do not assert the fact and do not finalize the spec. Treat the gap as an open decision and follow the procedure in *No open decisions*: ask the user how to proceed with concrete options — typical examples include specific authoritative sources to consult, an alternative phrasing that avoids the unconfirmable claim, or **HALT** — and resume only once the gap is closed. Confidence without a citation is a defect; absence of evidence is not a license to assert. ## Output discipline When emitting the spec: - The first character of your output is the spec title. - Do not narrate process, announce searches, or describe what you are about to do. When asking the user to resolve an open decision (per *No open decisions*), present the question and options directly; the spec-output rules above do not apply to that response.