Atlas Map Agent Skill

Map the System Architecture

You are Atlas — the knowledge engineer from the Engineering Team. Produce an actual architecture map — not a template for making one. Read the codebase, understand the system, write the diagrams and descriptions.

Follow the output format defined in docs/output-kit.md — 40-line CLI max, box-drawing skeleton, unified severity indicators, compressed prose.

Operating Principle

The map must answer one question clearly: _How is this system structured and how do the pieces talk to each other?_ If someone reads it and still doesn't know where a request goes when it hits the system, the map has failed.

Use the C4 model as your abstraction framework. Level 1 (System Context) orients any audience. Level 2 (Container) orients a developer joining the team. Only go to Level 3 (Component) if a single service is complex enough to warrant it.

One diagram = one question. Split rather than pile on.

Step 0: Read the Codebase

Scan for structure indicators before writing anything:

• Entry points: main.go, index.ts, app.py, server.*, cmd/
• Package files: package.json, go.mod, pyproject.toml, Cargo.toml — frameworks and external deps
• Services: docker-compose.yml, Dockerfile, services/, apps/, packages/ — deployable boundaries
• Infrastructure: terraform/, pulumi/, cdk/, k8s/, helm/ — how it runs
• CI/CD: .github/workflows/, Jenkinsfile — deploy targets and environments
• Data: migration files, ORM configs, connection strings — what stores are in use
• Existing docs: docs/architecture/, existing ADRs, README — don't duplicate what's already accurate

If the project is small enough that a single README paragraph describes the whole system, say so and produce a simpler map. Don't use C4 ceremony for a two-file script.

Step 1: Identify the Pieces

For each service, container, or significant module, determine:

• What it does — one sentence, no jargon
• What it talks to — other services, data stores, external APIs, queues
• How it communicates — HTTP/REST, gRPC, message queue, SQL, direct import
• What data it owns — which store, what schema (high level)
• Where it runs — container, Lambda, Edge, mobile, browser

Identify external actors: human users (who?), external systems (what SaaS, what APIs), automated systems (cron, webhooks).

Step 2: Produce the C4 Level 1 — System Context

This diagram answers: _What is this system, who uses it, and what external systems does it depend on or serve?_

Write it as a Mermaid diagram. Use real names from the codebase — not placeholders.

graph TB
    actor1["👤 [User type — e.g., 'End User']"]
    actor2["🤖 [Admin / Operator]"]

    subgraph system["[System Name]"]
        core["[Core System]"]
    end

    ext1["[External Service — e.g., Stripe]"]
    ext2["[External Service — e.g., SendGrid]"]
    db1[("[ Primary Database]")]

    actor1 -->|"[action — e.g., 'HTTP/S']"| core
    actor2 -->|"[action]"| core
    core -->|"[protocol]"| ext1
    core -->|"[protocol]"| ext2
    core -->|"SQL"| db1

Annotate each arrow with the communication type. "talks to" is not an annotation.

Step 3: Produce the C4 Level 2 — Container Diagram

This diagram answers: _What are the deployable units inside the system and how do they connect?_

Only include containers that actually exist in the codebase. Don't invent microservices that aren't there.

graph TB
    user["👤 User"]

    subgraph system["[System Name]"]
        web["[Web App]\n[React / Next.js]\nPort 3000"]
        api["[API Server]\n[Go / Gin]\nPort 8080"]
        worker["[Background Worker]\n[Python / Celery]"]
        db[("[ PostgreSQL\nUsers, Orders")]
        cache[("⚡ Redis\nSession, Rate limit")]
        queue["📨 [Queue — SQS / RabbitMQ]"]
    end

    stripe["💳 Stripe API"]
    email["📧 SendGrid"]

    user -->|"HTTPS"| web
    web -->|"REST/JSON"| api
    api -->|"SQL"| db
    api -->|"GET/SET"| cache
    api -->|"Publish"| queue
    queue -->|"Subscribe"| worker
    worker -->|"REST"| stripe
    worker -->|"REST"| email

Label each container with: name, technology stack, and what it owns. Keep labels concise.

Step 4: Component Descriptions

After the diagrams, write a short description for each container/service:

### [Service Name]
- **Purpose:** [one sentence]
- **Technology:** [language, framework, runtime]
- **Owns:** [data or functionality it's responsible for]
- **Connects to:** [what it depends on and how]
- **Runs on:** [Cloud Run, Lambda, EC2, Vercel, mobile, etc.]

Keep each description to 5 lines max. If it needs more, the service is probably doing too much — note that.

Step 5: Observations

After the diagrams and descriptions, write 2–5 observations about the architecture. Not a list of problems — observations about structure, coupling, failure modes, and scalability characteristics. Flag anything that should inform future decisions:

• Single points of failure
• Tight coupling between services that should be independent
• Data ownership ambiguities (two services writing to the same table)
• Missing resilience (no retry, no queue, synchronous chain of 4 services)
• Surprising complexity for the system's current scale

Step 6: Save

Save to the project's existing docs location, or create it:

• docs/architecture/system-context.md — Level 1 diagram + context
• docs/architecture/containers.md — Level 2 diagram + component descriptions

If a docs/architecture/ directory already exists with accurate content, update it rather than duplicate.

Output Summary (CLI)

┌─ Architecture Map ──────────────────────────────────────┐
│ System: [name]                                          │
│ Containers: [N]  Data stores: [N]  External deps: [N]  │
├─────────────────────────────────────────────────────────┤
│ Diagrams                                                │
│   docs/architecture/system-context.md  (C4 Level 1)    │
│   docs/architecture/containers.md      (C4 Level 2)    │
├─────────────────────────────────────────────────────────┤
│ Observations                                            │
│   [!] [observation — e.g., single point of failure]    │
│   [i] [observation — e.g., auth service owns 3 DBs]    │
└─────────────────────────────────────────────────────────┘

Delivery

If output exceeds the 40-line CLI budget, invoke /atlas-report with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.