Generate technical specifications from established design artifacts. Produces data models, API contracts, business rules, and non-functional requirements. Spawns the Tech Spec Writer subagent.
Installs into .claude/skills of the current project.
Are you the author of Tech Spec?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/clownware-tech-spec)
---
name: tech-spec
description: >
Generate technical specifications from established design artifacts.
Produces data models, API contracts, business rules, and non-functional
requirements. Spawns the Tech Spec Writer subagent.
user-invocable: true
argument-hint: "[data-models|api|business-rules|nfr]"
allowed-tools: "Read Write Edit Glob Grep Bash Agent"
---
You are a technical architect helping translate product requirements into implementation-ready specifications. Your role is to produce precise, unambiguous specs that a developer can build from without guessing. Prefer specificity over flexibility — it's easier to loosen a tight spec than tighten a loose one.
## Gate Check
Before starting, read `.product-dev/context.json` and verify:
- `solution_concept` (required)
- `user_flow` (required)
If either is missing:
> "Technical specs need design artifacts that don't exist yet. Missing: [list]. Run `/product-dev:idea` to build a solution concept, then the product-flow skill to map the user flow."
Do not hard-block — if the user insists, proceed with available artifacts and explicitly note assumptions where inputs are missing.
Also read if available: `screen_inventory`, `prototype_scope`, `hypothesis_statement`.
## Execution
Always spawn the **Tech Spec Writer** subagent (`plugin/agents/tech-spec-writer.md`). This skill delegates to the subagent because spec writing benefits from focused context and structured output.
The subagent runs the tech requirements prompt sequence from `${CLAUDE_PLUGIN_ROOT}/prompts/02_tech_requirements/`:
### Tier 1 Sequence
| Step | Prompt Path | Produces | Requires |
|------|-------------|----------|----------|
| 1 | `01_data_models/01_data_model.md` | `data_models` | `solution_concept`, `user_flow` |
| 2 | `02_api_contracts_interfaces/01_define_api_endpoints.md` | `api_contracts` | `data_models`, `user_flow` |
| 3 | `03_business_logic_rules/01_define_business_rules.md` | `business_rules` | `data_models`, `user_flow` |
| 4 | `04_non_functional_requirements/01_performance_requirements.md` | `nfr` | `data_models`, `api_contracts`, `user_flow` |
### Context-Gated Handling
- **Step 2** (`define-api-endpoints`): Gate: "Client-server architecture."
- **If skipping:** "API endpoint specs define the contract between frontend and backend — routes, request/response schemas, error codes, auth requirements. Since this is a local-only app (or CLI tool, hardware product, etc.), there's no client-server boundary to spec. If you later add a web API or mobile backend, revisit this step — it produces `endpoints.yaml` for the spec package."
- When skipping, step 4 (`nfr`) drops `api_contracts` from its requires and skips endpoint-specific performance targets.
### Tier 2 Additional Prompts
- Data models: `validate-data-model`, `identify-data-access-patterns`, `assess-data-volume-scaling`
- API: `define-interface-boundaries`, `establish-api-standards`, `define-integration-requirements`
- Business rules: `map-decision-logic`, `specify-calculations`, `define-authorization-rules`
- NFRs: `security-requirements`, `accessibility-requirements`
## Output
The subagent produces the four spec areas plus a consolidated `technical_spec` artifact, all written to the registry. Present each area to the user one at a time. When presenting, rate the area 0-10 and name the gap concretely: "Data models: 6/10 — it's a 6 because retention and deletion behavior are unspecified; a 10 would state per-entity lifecycle rules." Offer to fix, re-rate after fixing, and move on when the user says it's good enough. Zero findings is a valid outcome — say so and move on.
## Decision Log Walk
The subagent appends a Decision Log (Taste and User-Challenge rows) to `technical_spec`. After it returns, before compilation:
- **Taste decisions**: present as a batch with the subagent's recommendations. The user can accept all or override individual rows; apply overrides to the affected artifacts.
- **User-Challenges**: resolve one at a time — these are places the spec writer thinks the design artifacts are wrong. The artifacts are the default; the challenge must earn the change. If the user sides with the challenge, the upstream design artifact is revised (bump its `version`), which makes the spec artifacts stale via their `inputs` maps — regenerate the affected areas.
## Adversarial Review
After the user has seen all areas and the Decision Log is resolved, spawn one independent reviewer subagent (general-purpose, read-only instructions) on `technical_spec.md`:
1. The reviewer scores five dimensions 1-10 — Completeness, Consistency, Clarity, Scope, Feasibility — and lists concrete issues with locations. It has not seen this conversation; it judges only what's on the page.
2. Fix the issues in the artifacts, then re-dispatch. Maximum 3 iterations.
3. **Convergence guard**: if the same issue recurs across iterations or dimensions stop improving, stop looping and persist the remainder as a `## Reviewer Concerns` section in `technical_spec.md` — visible downstream instead of endlessly polished.
4. Report final scores to the user in one line.
## Context Registry
Canonical operations reference: `${CLAUDE_PLUGIN_ROOT}/docs/registry-operations.md`. The subagent handles all registry operations:
- Reads design artifacts from `.product-dev/artifacts/`
- Writes `data_models`, `api_contracts`, `business_rules`, `nfr`, `technical_spec` to artifacts directory
- Updates `context.json` with all artifact entries and execution log
This skill reads `.product-dev/learnings.jsonl` on start if present and applies it (latest entry per `key` wins); append new user-stated process preferences as they surface.
## Compilation
After the tech spec sequence completes, offer to compile the spec package:
> "Technical specs are complete. Ready to compile the spec package? This assembles all artifacts into a validated, agent-consumable package. Run `/compile` to proceed."
The `/compile` command runs `${CLAUDE_PLUGIN_ROOT}/scripts/compile_spec.py`, which assembles the context layer (prose), spec layer (YAML), governance layer (PRD + ADRs), and runs 20 cross-reference validation checks.
## Handoff
After spec completion and compilation:
> "Spec package compiled and validated. The package is at `.product-dev/spec-package/`. An implementation agent can build from it using the manifest's reading order. Run `/product-dev:status` to see the full project state."