Generates a full design artifact set from a spec.md (primary source) and optional architecture.md (secondary source). Produces three outputs: architecture.md (system architecture), data-model.md (persistent/domain data model), and API_IMPLEMENTATION_GUIDE.md (API contract). Use this skill whenever a user provides a spec, architecture, or requirements document and asks for a design, detailed design, solution design, module design, or implementation design. Also trigger when the user says "convert architecture to design", "create design from architecture", "write a design document", "I need a design doc before we start coding", "break down this architecture for engineers", or "generate design artifacts". If the user uploads or pastes an architecture or spec document and wants to move toward implementation, this skill should always be used — even if they don't use the exact words "design document".
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Generates a full design artifact set from a spec.md (primary source) and optional architecture.md (secondary source). Produces three outputs: architecture.md (system architecture), data-model.md (persistent/domain data model), and API_IMPLEMENTATION_GUIDE.md (API contract). Use this skill whenever a user provides a spec, architecture, or requirements document and asks for a design, detailed design, solution design, module design, or implementation design. Also trigger when the user says "convert architecture to design", "create design from architecture", "write a design document", "I need a design doc before we start coding", "break down this architecture for engineers", or "generate design artifacts". If the user uploads or pastes an architecture or spec document and wants to move toward implementation, this skill should always be used — even if they don't use the exact words "design document".
Architecture-to-Design Skill
Purpose
Generate a complete, implementation-ready design artifact set from a specification document.
The output bridges the gap between product requirements and engineering execution — detailed enough
for task breakdown and coding, but not so low-level that it becomes code or pseudocode.
Artifact Set
This skill produces up to four files per invocation:
Primary source of truth for requirements, entities, states, workflows, roles, integrations
2
architecture.md (if it exists)
Secondary source for technology choices, layer decisions, component grouping
3
Existing repo code and config
Validation and refinement — confirm what is already implemented; do not contradict working code
4
Sample files provided by user
Style reference only — formatting, tone, abstraction level; never copy sample content
If spec.md and architecture.md disagree, prefer spec.md.
If spec.md and implemented code disagree, note the discrepancy in the design as an open question.
Workflow
Step 0: Codebase Grounding Pass (MANDATORY)
Before writing any design artifact, scan the source spec/architecture for every claim about existing code and verify each against the real repository. This single step prevents the vast majority of review findings.
Required verifications:
Every method name referenced in spec/architecture — grep it. Use file.java:line anchors in the generated design, not just method names
Every claim like "the existing X already supports Y" — verify or tag [UNVERIFIED]
Every annotation (@PreAuthorize, @Transactional, etc.) — confirm the real codebase uses the same style, or use the actual pattern (e.g., imperative validation helpers)
Every upstream assumption — the spec may be wrong. Re-verify; do not inherit blindly (Rule 6 in grounding-rules)
Every signature you plan to cite — confirm parameter names, parameter counts, return types
Design-specific requirement: Because the design document carries implementation-facing detail (file paths, class names, method signatures), every reference to an existing file MUST include a real line anchor like ReleaseFlowService.java:172. If you cannot produce a line anchor, tag the reference [UNVERIFIED].
See ../_shared/grounding-rules.md (Rules 1, 2, 6) for the full protocol.
Step 1: Locate and Ingest Sources
Check for source documents in this order:
Inline content provided in the conversation
Files in the current workspace: look for docs/**/spec.md, docs/**/architecture.md
A path explicitly provided by the user
Read the spec fully. Read the architecture if it exists. Optionally scan the repo for implemented
controllers, entities, DTOs, and config to validate design decisions.
If no spec or architecture is found, ask the user to provide one before proceeding.
Step 2: Analyze the Specification
Extract and mentally model:
System purpose and goals — what problem does this system solve?
Primary users / actors — who interacts with the system, and how?
Core capability domains — what are the major functional areas?
Entities and data model — what are the key entities, their attributes, relationships, and state models?
Workflows and lifecycle stages — what are the key end-to-end flows? What states does each entity move through?
API surface — what endpoints, operations, and contracts does the system expose?
External systems and integrations — what does this system depend on or call out to?
Non-functional requirements — security, auth, audit, performance, reliability — where the spec supports them
Gaps and ambiguities — anything the spec does not address that materially affects design
Step 3: Plan the Design Scope
Before writing, determine:
Which modules need detailed design?
Which workflows are most complex or implementation-sensitive?
Where are the integration boundaries?
What is explicitly in scope vs. out of scope?
What must be inferred vs. what is explicitly stated?
Does the system expose an API? (determines whether artifact #4 is generated)
Does the system persist data? (determines whether artifact #3 is generated)
Step 4: Generate Artifact 1 — System Architecture
Produce docs/04-architecture/architecture.md.
Content:
Technology stack (only what spec or architecture explicitly states)
Constraints and assumptions
High-level architecture diagram (plain-text ASCII block diagram in a fenced code block — not Mermaid)
Data architecture: conceptual entities table, field-to-domain mapping, persistence notes
4–8 primary boxes, grouped by layer (Users → Frontend → API → Core Domain → DB + External Systems)
Do NOT use Mermaid unless the user explicitly requests it
Readable in 10 seconds — documentation block diagram, not implementation map
Step 5: Generate Artifact 2 — Design Document
Produce docs/05-design/design.md.
Content — use this structure:
# Detailed Design: [System Name]
## Overview
## Source Architecture
## Design Assumptions
## Design Scope
## Module Design — one subsection per major module
## API / Interface Design — endpoints, inputs/outputs, validation, errors
## Data Design — entities, relationships, state transitions
## UI / User Flow Design — views, actions, feedback (omit if no UI)
## Workflow / Execution Design — step-by-step flows, ordering, failure handling
## Integration Design — per external system: pattern, credentials, retry
## Security / Audit / Reliability Design
## Validation and Error Handling
## Testing Considerations
## Risks / Design Tradeoffs
## Open Questions
Key generation rules:
Derive from spec — do not invent capabilities not supported by it
If inferred, label as [Assumption]
Surface gaps in Open Questions, do not silently resolve
Keep language concise, professional, engineering-friendly
Do not write code, SQL DDL, or exact payload schemas unless explicitly requested
Make state transitions, validation rules, and error handling explicit
For workflow/DevOps platforms: always design configuration handling, execution state, audit history, and integration behavior
Do not invent frameworks, databases, or infrastructure unless stated in spec; if inferred, label [Assumption]
Step 6: Generate Artifact 3 — Data Model (if applicable)
Produce docs/04-architecture/data-model.md if the spec defines persistent entities.
When to generate: The spec has a Data Model, Entity, or State Model section with at least one entity.
Content source: Derive from spec entity definitions, field lists, relationships, and state models. Cross-reference with architecture data section and existing repo entities if available.
Required structure:
# Data Model: [System Name]
## Overview — 1–2 sentences on what is persisted and why
## Entity Relationship Diagram — ASCII ER diagram showing cardinality
## Entity Definitions — one subsection per entity:
- Table name
- Columns: name, type, nullable, description
- Primary key, foreign keys, unique constraints
- Indexes (where spec implies query patterns)
## State Models — per stateful entity:
- Valid states (enum values)
- Valid transitions (from → to, trigger)
- ASCII state diagram
## Configuration Entities — config keys, types, validation rules
## Audit Entities — audit record structure, action types, immutability rules
## Field Mapping — source-to-entity mapping table (e.g., Excel template → domain)
Rules:
Use column types at the logical level (String, Integer, Timestamp, JSON/CLOB, Boolean) — not vendor-specific DDL
Include all entities from the spec, including supporting entities (audit, config, execution history)
Mark any entity or field not in the spec as [Assumption]
If the spec defines a field mapping (e.g., template fields → entity columns), include it verbatim
ER diagram rules:
Plain-text ASCII using box-drawing characters
Show entity names and cardinality (1:N, N:1, 1:1)
Group related entities visually
Skip this artifact if the spec has no persistent entities (e.g., a pure proxy or stateless service).
Produce docs/05-design/contracts/API_IMPLEMENTATION_GUIDE.md if the design defines HTTP/REST endpoints.
When to generate: The design includes an API / Interface Design section with at least one endpoint.
Content source: Derive from the design document produced in Step 5. Cross-reference with spec functional requirements and existing repo controllers if available. Do not invent endpoints or fields not in the design.
Required structure:
# [System Name] — API Implementation Guide
Date / Version / Base Path / Backend stack / Auth model
## Overview — 1–2 sentences
## Authentication — session/token model, auth chain, roles table, stub users (if any)
## Error Response Format — standard error body shape, error code table
## API Endpoints Summary — one table per functional domain:
Operation | Method | Endpoint | Auth
## Endpoint Reference — per endpoint:
- Purpose
- Request params (path, query, body with field/type/required)
- Example request body (JSON)
- Example response body (JSON)
- Validation rules
- Error cases (status + when)
- Side effects (persistence, audit, integrations)
## State Reference — ASCII summary of key status models and transitions
## Concurrency — optimistic locking or conflict handling
## Integration Dependencies — external systems, required config, protocol, timeouts
Style rules:
Implementation-facing, concise, structured
Prefer concrete endpoint descriptions over generic prose
Match documentation tone of the design document
No meta commentary
Skip this artifact if the design has no API surface (e.g., a pure library or batch system).
Step 7.5: Pre-Ship Self-Review (MANDATORY)
Before delivering any artifact, run the shared Pre-Ship Checklist from ../_shared/grounding-rules.md across every generated file:
F1 — Every method/class/file/annotation reference is grep-verified and anchored with file.ext:line, or tagged [UNVERIFIED]
F2 — Every "the existing X already supports Y" claim is verified against real code
F3 — The architecture doc contains no design/task-level content (no code bodies, no precise method signatures, no LOC estimates, no test file names). The design doc contains no task-level content (no task IDs, no PR decomposition, no phase gates, no owner assignments)
F4 — Scope / Assumptions / Design Boundaries / Testing Considerations do not contradict each other. In particular: test expectations must match the rules they are testing, and Module descriptions must match the [Out of Scope] list
F5 — Every new rule (regex, ordering, sorting, algorithm, enum, routing) has been traced against at least 3 edge cases, including one case where the rule is intentionally NOT supposed to fire
F6 — No "implementation will decide" / "grep and choose later" language. Every implementation-impacting decision (DTO field order, method parameter shape, fallback behavior, boundary response codes) is committed
F7 — Every claim inherited from the architecture or spec has been re-verified, not carried forward blindly. Known phantom-inheritance hot spots: claims about existing service method names, repository method names, authorization style (@PreAuthorize vs imperative), load patterns (findByIdWithFullHierarchy vs getById + findRequestsForFlow)
If any box is unchecked, fix the gap before delivering.
Step 8: Deliver All Artifacts
Write each artifact to its designated path. Summarize what was generated:
