Architecture spec

Skill: Architecture Spec (Post-Implementation)

Type: Execution

Purpose

Given implemented changes (diff / changed files), generate an architecture/design document with appropriate depth.

The depth is selected automatically based on:

• Risk signals
• Layer spread
• Change magnitude
• Sensitive areas (auth, infra, migration, etc.)

---

When to Use

• After completing a feature or significant code change
• Before merging a PR that touches multiple layers or sensitive areas
• When stakeholders need a design record for changes already implemented
• After a hotfix or incident-driven change that requires documentation

---

When NOT to Use

• During initial design/planning (use a design doc skill instead)
• Documentation-only or comment-only changes
• Auto-generated code changes (lock files, migration snapshots)
• Trivial single-file changes with no risk signals (e.g., typo fix)

---

Inputs Required

Do not run this skill without:

• changed_files (list of modified files)
• diff_summary (LOC count or file count)
• feature_name (short name for the change)

Optional but recommended:

• diff_snippets (key portions of the diff)
• repo_context (architecture overview, dependency map)

Without asking the user (unless unavailable), gather these inputs directly from the repository.

---

Output Format

1. Risk Evaluation Summary (score breakdown, selected level)
2. Architecture/Design Document (level-appropriate Markdown)
3. ADR section (Level C only)
4. Notion page URL (if Notion integration available)

---

Procedure

Step 1 — Evaluate Risk

→ Detailed scoring criteria: subskills/diff-risk-evaluator.md (load only if the summary below is insufficient for scoring)

Analyze changed files and calculate a deterministic risk score across three dimensions:

Dimension	What it measures
Path-Based Risk	Sensitive area keywords in file paths (+3 to +4 per match)
Layer Spread	Number of architectural layers touched (+1 to +5)
Change Magnitude	Lines of code changed (+1 to +6)

total_score = path_score + layer_score + magnitude_score

Step 2 — Select Documentation Level

Total Score	Level
0–6	A (Lightweight)
7–13	B (Standard)
14+	C (Architecture-Level)

Hard Rules:

• Auth + multi-layer → minimum B
• Infra change → minimum B
• Migration + medium change → minimum B
• Financial impact → C
• Global middleware + high magnitude → C

If unsure, prefer B over A.

Step 3 — Generate Document

→ Load and follow subskills/notion-spec-generator.md at this point (not before).

Generate a Notion-ready Markdown document for the selected level:

• Level A — Overview, What Changed, Simple Flow, Decisions, Test Notes
• Level B — Level A + Architecture, Sequence Diagram, API Spec, Edge Cases, Security Notes, Operational Notes, Future Improvements
• Level C — Level B + Threat Model, Failure Flow, Rollback Plan, Observability Plan, ADR-style Decisions

For Level C only, additionally load and follow subskills/adr-generator.md to produce a formal ADR section. Do NOT load this subskill for Level A or B.

All levels must include: Title, Metadata table, TL;DR, Changed files summary.

Step 4 — Publish to Notion (Optional)

ACTIVATION: This step runs only if Notion MCP integration is available and the user requests publishing. If skipped, do NOT load the subskill file.

→ Load and follow subskills/notion-page-publisher.md only when this step is activated.

If Notion MCP integration is available, persist the generated spec to Notion:

• Create or update a page in the target database
• Map properties (Level, Risk Score, Feature, Status)
• For Level C, attach ADR as a child page or appended section
• If a Draft page with the same feature name exists, update instead of creating a new one
• Return the Notion page URL

---

Quality Bar

The document must answer:

• What changed?
• Why?
• How does it work?
• Where can it fail?
• How is it monitored?
• How do we roll back?
• Is the document visually scannable? (emoji headings, tables, diagrams, dividers)

---

Guardrails

• Deterministic scoring first; keyword detection only nudges the level.
• Never under-document security or financial changes.
• Prefer a safer (higher) level if signals are incomplete.
• Do not invent architecture details not present in the code.
• Explicitly state assumptions when context is incomplete.
• If inputs are insufficient to evaluate risk, ask for clarification before proceeding.
• Do not skip the risk evaluation step and jump directly to document generation.

---

Failure Patterns

Common bad outputs:

• Selecting Level A for multi-layer changes that touch auth or infra
• Generating a document without running the risk scoring procedure
• Producing generic architecture descriptions not tied to the actual diff
• Missing the ADR section for Level C changes
• Inflating risk scores by double-counting the same file in multiple categories
• Skipping the Quality Bar questions (especially "Where can it fail?" and "How do we roll back?")

---

Example 1 (Minimal Context)

Input:

feature_name: "Add rate limiting to public API" changed_files: middleware/rate-limiter.ts, config/rate-limit.ts, routes/api.ts diff_summary: 120 LOC

Output:

1. Risk Evaluation:
   • Path score: +3 (middleware) + +3 (config) = 6
   • Layer spread: +3 (2 layers: middleware, routes)
   • Magnitude: +1 (<150 LOC)
   • Total: 10 → Level B
2. Document: Standard spec with Overview, Architecture (middleware chain diagram), Sequence Diagram (request → rate check → pass/reject), API Spec (429 response), Edge Cases (distributed rate limiting gaps), Security Notes (bypass vectors), Operational Notes (Redis dependency)

---

Example 2 (Realistic Scenario)

Input:

feature_name: "Migrate user auth from session to JWT" changed_files: auth/jwt-provider.ts, auth/session-provider.ts (deleted), middleware/auth.ts, config/auth.ts, migrations/20250220_drop_sessions.sql, routes/login.ts, routes/logout.ts, services/user.ts, tests/auth.test.ts diff_summary: 850 LOC

Output:

1. Risk Evaluation:
   • Path score: +4 (auth) + +3 (middleware) + +3 (config) + +3 (migration) = 13
   • Layer spread: +5 (4+ layers: auth, middleware, config, routes, services)
   • Magnitude: +4 (500–1500 LOC)
   • Total: 22 → Level C
   • Hard rule applied: Auth + multi-layer → minimum B (already exceeded)
2. Document: Full architecture spec with Threat Model (token theft, replay attacks), Failure Flow (JWT validation failure paths), Rollback Plan (session table restoration, dual-auth transition period), Observability Plan (auth failure rate metric, token expiry distribution)
3. ADR: "Migrate from server-side sessions to JWT" — options considered (session + Redis vs JWT vs JWT + refresh token), decision rationale, consequences (stateless scaling benefit vs token revocation complexity)

---

Notes

FAST MODE (only if explicitly requested):

• Always use Level A regardless of risk score
• Skip Step 4 (Notion publish)
• Risk evaluation still runs for the metadata record

---

This skill delegates detailed work to four subskills:

• diff-risk-evaluator — risk scoring only (no document generation)
• notion-spec-generator — Markdown document generation per level
• adr-generator — ADR section for Level C changes
• notion-page-publisher — Persist spec to Notion (optional, requires Notion MCP)