atlas
This skill turns a stated need into a real systems atlas by SCANNING your
actual sources — Azure subscriptions (via read-only az), git repos, and local
directories — and process/data mining them, THEN enriching the result against the
Atlas knowledge graph. It is the brain of the atlas plugin. The scan is
PRIMARY; the graph is SECONDARY. For non-trivial runs it delegates orchestration
to babysitter:babysit using an atlas-specific .a5c process; for simple
lookups it queries the graph directly.
1. Scan-first, graph-second
The output you want is an evidence-backed inventory of your systems — e.g.
azure-inventory.json (every real resource id + RG from az),
workspace-inventory.json (real repo/dir scan), processes.json (real mined
CI/CD/IaC/.a5c processes), and a cross-linked SYSTEMS-ATLAS.md. Every item must
cite its REAL source. Generic catalog nodes are NOT the deliverable.
- Primary — scan the user's real sources. Use
Bash to run READ-ONLY scans:
az (account/group/resource list + per-service list/show) for Azure;
git + filesystem (Read/Glob) for repos and directories. NEVER invent
resource ids, regions, SKUs, or file paths — if you didn't observe it in real
output, it does not go in the atlas. Only scan the sources named in the need
(scoping, not a fallback).
- Secondary — the Atlas knowledge graph. Atlas is a knowledge graph of
agents, processes, data models, capabilities, workflows, and wiki pages reached
through the
mcp__atlas__atlas_public_* MCP tools (server URL overridable via
ATLAS_MCP_URL). Use it ONLY to add best-practice / comparison context for the
real systems you found — never as the primary content, never to pad the atlas
with generic nodes. See the atlas-graph-query skill for the tool surface.
2. When to use
| Trigger phrase | Command |
|---|
| scan/inventory my real systems (azure + repos + dirs), map them | /atlas:discover |
| mine the real processes in my repos/cloud (CI/CD, IaC, .a5c, cron) | /atlas:mine-processes |
| mine the real data stores/models in my cloud + repos | /atlas:mine-data |
| collect the real constraints/gotchas of my scanned systems | /atlas:collect-nuances |
3. The need → real atlas pipeline (core method)
- Parse sources — interpret the stated need into concrete SOURCES: Azure
subscription(s), git repos, local directories, URLs, plus the output dir. If
the sources are genuinely ambiguous, run a short interview
(
AskUserQuestion). Per repo policy, interview ONLY when truly unclear.
- Scan cloud (primary) — for each Azure source, run read-only
az and write
a real cloud inventory citing resource ids/RGs. Skip cleanly (record a reason)
if no cloud source is in scope — only scan what's named.
- Scan local (primary) — for each repo/dir, scan the filesystem + git
(structure, submodules, manifests, languages, services, IaC) and write a real
inventory citing real paths.
- Enrich (secondary) — map the discovered real systems against the Atlas
graph for comparison context. Clearly secondary; never the headline.
- Synthesize — assemble a real, cross-linked layered atlas (components /
processes / data / integrations / nuances) where EVERY item cites its real
source, like
SYSTEMS-ATLAS.md, plus a machine mirror.
- Converge (TDD) — each phase asserts its own checkable outputs before
proceeding (see the atlas processes), iterating until the assertions pass.
4. How to delegate
For any non-trivial run, hand off to babysitter:babysit (via the Skill tool)
naming the matching atlas process:
/atlas:discover → atlas-systems-discovery/atlas:mine-processes → atlas-process-mining/atlas:mine-data → atlas-data-mining/atlas:collect-nuances → atlas-collect-nuances
Do not hand-roll orchestration when a process exists.
5. Guardrails
- No fallbacks (repo rule). Skipping an out-of-scope source class (e.g. no cloud
named) is correct scoping and must be recorded with a reason — it is NOT a
silent fallback to the public graph. If you find yourself writing a real
fallback, stop and fix the root cause.
- Scan-first: every system/item in the atlas MUST cite a REAL source (an
az
resource id / RG, or a file path). Never invent resource ids, regions, SKUs, or
file paths. Never invent graph node ids either — only reference ids returned by
the Atlas tools, and keep graph content strictly secondary.
- Read-only scanning only:
az read verbs, git status/remote/log, filesystem
reads. Never run mutating cloud/git/fs commands and never read secret values.
- Keep breakpoints sparse; use them only when the sources to scan are genuinely
ambiguous.
- The real scanning is done BY the agent via its
Bash tool inside the agent
task prompt. Do not emit kind: 'shell' subtasks unless the user explicitly
asks for a shell-oriented workflow.
a5c-ai· 8d agoSandbox