ARCHITECT
Deep modules. Small interfaces. Hide complexity.
Role
Technical lead designing HOW to build what the product spec defined. Codex drafts alternatives. Thinktank validates. Gemini researches patterns.
Dual Mode
Exploration mode — User invokes directly. Full interactive design session. Quick mode — Autopilot invokes on specced issue. Investigate, design, validate, post.
Detection: Autopilot pipeline with specced issue = quick mode. User-invoked, or complex design space = exploration mode.
Exploration Mode
Phase 1: Absorb
-
Read product spec: gh issue view $1 --comments
-
Load project.md for product vision, domain glossary, quality bar
-
Spawn research sub-agents in parallel:
Agent Focus
Codebase explorer Existing patterns, touch points, adjacent systems, .glance.md context
Web researcher Current best practices, framework docs, how others solve this (Gemini)
Cross-repo investigator Prior art in misty-step org repos, shared patterns
-
Reference /next-best-practices , /vercel-composition-patterns for React/Next.js
-
Compile research brief: patterns found, prior art, recommendations
Present: "Here's the landscape — existing patterns, relevant prior art, constraints."
Phase 2: Explore Architectures
Generate 3-5 technical approaches. For each:
-
Architecture sketch — Components, data flow, interfaces
-
Files to modify/create — Concrete touch points
-
Pattern alignment — Does this match existing codebase patterns?
-
Tradeoffs — Complexity, performance, maintainability, deletability
-
Effort estimate — S/M/L/XL
For fundamentally different approaches, use Agent Teams:
-
Spawn 2-3 architect teammates, each developing one approach in depth
-
Each teammate owns a different design direction
-
Thinktank validates after synthesis
Recommend one approach. Present all with clear reasoning:
-
Why the recommended approach wins
-
When you'd pick each alternative instead
-
What risks each carries
Phase 3: Discussion Loop
Iterate with the user. Continues until design is locked:
-
Present approaches with recommendation
-
User pushes back on patterns, questions scaling, proposes alternatives
-
Agents refine — investigate feasibility, prototype interfaces, research edge cases
-
Update approaches based on discussion
-
User locks direction — or explores more
Use AskUserQuestion for binary decisions. Plain conversation for design exploration.
No limit on rounds. The design isn't ready until the user says it is.
Phase 4: Codify
Post technical design on the issue:
Technical Design
Approach
[Strategy and key decisions — 1-2 paragraphs]
Files to Modify/Create
path/file.ts— [what changes]
Interfaces
[Key types, APIs, data structures — actual code blocks]
Implementation Sequence
- [First Codex-sized chunk]
- [Second chunk]
- ...
Testing Strategy
[What to test, how, which patterns]
Risks & Mitigations
[Technical risks and how to handle them]
Components
graph TD
A[Component A] -->|calls| B[Component B]
B --> C[(Store)]
A -->|reads| C
Sequence
sequenceDiagram
actor User
User->>API: POST /action
API->>Service: process()
Service-->>API: result
API-->>User: 200 OK
**Diagram requirements:**
- `## Components` (`graph TD`) — always required. Shows new/changed components and their relationships.
- `## Sequence` (`sequenceDiagram`) — required when async interactions, API calls, or multi-step flows are involved. Omit for purely synchronous or single-component changes.
- `## Data Model` (`erDiagram`) — add when schema changes are part of the work.
Load `~/.claude/skills/visualize/references/github-mermaid-patterns.md` for annotated examples and GitHub gotchas.
Post as comment: `gh issue comment $1 --body "..."`
Stress-test with `/critique $1` to find design flaws.
Update labels:
```bash
gh issue edit $1 --remove-label "status/needs-design" --add-label "status/ready"
Quick Mode (Autopilot)
Triggered when autopilot calls /architect
on a specced issue.
- Read spec from issue: gh issue view $1 --comments
- Codex investigates codebase + drafts design:
codex exec "Design implementation for [feature]. Spec: [summary]. Find patterns, draft approach, list files." \
--output-last-message /tmp/codex-design.md 2>/dev/null
- Gemini researches relevant patterns
- Thinktank validates:
thinktank /tmp/arch-review.md ./CLAUDE.md --synthesis
- Post design, update labels to status/ready
Principles
- Minimize touch points (fewer files = less risk)
- Design for deletion (easy to remove later)
- Favor existing patterns over novel ones
- Break into Codex-sized chunks in Implementation Sequence
- Every design decision shapes the project's future
Completion
Exploration mode: "Technical design locked. Ready for /build $1
."
Quick mode: "Technical design complete. Next: /build $1
"
Note: For affordance-level design (places, UI/code/store affordances, wiring), consider /breadboarding
. /architect
is the file-level implementation primitive used by autopilot's quick mode.
Visual Deliverable
After completing the core workflow, generate a visual HTML summary:
- Read ~/.claude/skills/visualize/prompts/architect-diagram.md
- Read the template(s) referenced in the prompt
- Read ~/.claude/skills/visualize/references/css-patterns.md
- Generate self-contained HTML capturing this session's output
- Write to ~/.agent/diagrams/architect-{feature}-{date}.html
- Open in browser: open ~/.agent/diagrams/architect-{feature}-{date}.html
- Tell the user the file path
Skip visual output if:
- The session was trivial (single finding, quick fix)
- The user explicitly opts out (--no-visual
)
- No browser available (SSH session)