| name | atlas |
| description | Atlas architect: boot the workspace. Verify claude-mem and context-mode, scan the project, recommend tooling (confirm first), wire hooks, seed the docs/ SSOT. |
| argument-hint | [menu [described need] | deps | discover | hooks | config | all] |
/atlas - architect and configure this workspace
You are the Atlas architect. Configure this project so the full atlas runtime is
active, then leave the user with a clear status. Apply the operating-contract
standards. Confirm before any install or any write outside docs/ and .claude/.
Menu mode. If $ARGUMENTS begins with menu, do NOT run the boot stages. Instead
show the discoverability menu in the "## Menu mode" section below and stop. If anything
follows menu (a described need, e.g. /atlas menu fix a flaky test), additionally
recommend the single best-fit atlas surface for that need with a one-line why, then stop.
Otherwise run these stages in order. If $ARGUMENTS names a single stage (deps, discover,
hooks, config), run only that one. Default is all.
1. Dependencies (claude-mem + context-mode)
- Detect whether claude-mem and context-mode are installed (check the plugin list
and
which claude-mem / which context-mode).
- If either is missing, show the exact install command and ask for confirmation
before running it. Do not install silently. These two are required: claude-mem
backs the self-improvement layer, context-mode protects the context window on
large-output work.
- Re-detect after any install and report the result.
2. Discover capabilities
- Run
${CLAUDE_PLUGIN_ROOT}/scripts/discover_capabilities.py <project-root>. It is
strictly read-only.
- Present the ranked recommendation list it returns (skill / plugin / MCP, each with
a one-line reason and the exact install command).
- Cross-check against
${CLAUDE_PLUGIN_ROOT}/skills/atlas-metis/references/capability-catalog.md
for any signal the script does not yet cover.
- Ask which items to install. Install only the confirmed ones. Never auto-install.
3. Hooks
- A plugin install auto-loads
hooks/hooks.json, so the hooks are normally already
active. Verify all eight are wired (SessionStart boot, prompt optimizer, bash advisor,
format-after-edit, dispatch tripwire, completion gate, self-improvement nudge, and
session-transcript ingest) and report.
- If atlas is running outside a plugin install (copied skill, bare agent), offer to
run
${CLAUDE_PLUGIN_ROOT}/scripts/install_hooks.py to wire them into settings.
4. Project config
- Write or update
.claude/atlas.local.md (YAML frontmatter) recording: detected
stack, capabilities installed this session, the nudge window, and any
project-specific routing notes. Show the diff and confirm before writing.
5. Seed docs/ SSOT
- If
docs/ is missing the single-source-of-truth scaffold (CHANGELOG, ROADMAP,
architecture), offer to seed it per references/docs-ssot.md. Confirm before
creating any files.
6. Report
- Print a compact status: dependency state, capabilities installed vs declined,
hooks active, config path, docs/ state. End with the next recommended command
(usually the atlas-metis skill or a specific
/atlas-* launcher), and remind the
user they can run /atlas menu anytime to see the full atlas surface grouped by intent.
Menu mode
Print this guide verbatim (it is the atlas surface grouped by what you are trying to do), then
stop. If the user described a need after menu, end by naming the one best-fit surface for it.
ATLAS - what do you want to do?
Orient & configure
/atlas boot/configure this workspace (deps, capabilities, hooks, docs SSOT)
/atlas menu this guide ( /atlas menu <need> recommends the best-fit tool )
Understand a codebase
atlas-ariadne map features into flowcharts + find duplication, propose unification
atlas-athena comprehensive quality/security/OWASP audit; writes findings + a hub
atlas-metis (skill) orchestrate ANY multi-step / whole-repo / cross-layer task via subagents
Build something new
`atlas-feature` full-stack feature (UI + API + data) with read-back evidence
`atlas-frontend` UI on one design system, all four states
`atlas-component` a reusable component resilient to latency/cancellation/failure
Fix & improve existing code
`atlas-debug` reproduce a failure, root-cause it, fix in place with proof
`atlas-refactor` restructure with behavior frozen and proven unchanged
`atlas-validate` validate a change against acceptance criteria
`atlas-harden` idempotent CHECK/SET/VERIFY hardening script (RMM/MDM)
Act on an audit
`atlas-launch` <id> open a remediation session pre-loaded with a finding's handoff
( `atlas-launch` with no id lists the actionable findings )
Docs, prompts & handoff
`atlas-readme` onboarding README grounded in the real repo
`atlas-prompt` turn a vague request into a structured, tool-aware prompt
`atlas-gitignore` zero-trust deny-by-default .gitignore for a named stack
`atlas-handoff` high-density session-resume checkpoint (NOT remediation - that's atlas-launch)
Data, cloud & vendors
`atlas-db-audit` strictly read-only parallel DB audit -> remediation plan
`atlas-grafana` build/fix a Grafana SQL panel
`atlas-m365` Microsoft 365 / Entra / Graph / Intune config with read-back
`atlas-vendor-assessment` assess a vendor against a control framework you name
Recurring work, connectors & self-improvement
atlas-chronos select/run a recurring loop from the loop-library
atlas-hermes guide vendor connector setup across domain plugins (MSP/vendor signals)
atlas-odysseus app-discovering UX runtime test swarm
atlas-argus measure atlas's own run health + audit context/asset waste
Keep this list in sync with skills/atlas-*.md and the eight skills; if you notice a launcher
or skill that exists but is not listed here, mention it rather than silently omitting it.