Polaris Quickstart
Cheat sheet for scaffolding a new project. See USAGE.md for full details.
First-Time Setup (once)
-
git clone https://github.com/theyoungastronauts/polaris.git ~/prj/polaris && cd ~/prj/polaris -
./install.sh init— saves repo path, addspolarisalias -
source ~/.zshrc— pick up the alias -
polaris global --fresh— install global skills + developer defaults -
pip install axoniq— optional: Axon is a third-party code-graph tool (call graphs, impact analysis) that Polaris uses over MCP when present; everything works without it
Safe on an existing setup. Plain polaris global amends: CLAUDE.md content outside the Polaris markers is preserved, settings.json is merged (nothing of yours removed), and every file written is manifest-tracked so polaris uninstall removes exactly what was added. --fresh is the one flag that replaces your CLAUDE.md — it backs the old one up first. When in doubt, drop --fresh.
New Project
1. Initialize + Design
-
polaris new ~/prj/my-app— creates project, selects stacks, installs context -
cd ~/prj/my-app
Option A — brainstorm from scratch:
- Open Claude Code session
- "Let's brainstorm [your idea] using the brainstorming skill"
- The agent already knows your stack — no need to re-explain it
- Iterate until design is solid
- Design saved to
docs/plans/YYYY-MM-DD-<topic>-brainstorm.md
Option B — bring existing design work:
- Drop briefs, wireframes, sitemaps, flow diagrams into
docs/design/ - Open Claude Code session
- "Review my design materials using the design-intake agent"
- Answer clarifying questions as the agent distills the artifacts
- Design doc saved to
docs/plans/YYYY-MM-DD-<topic>-design.md
1b. Product Definition (optional, for UI features)
-
/prd— formalize requirements into a structured PRD -
/ux-spec— run 6 UX passes (mental model, IA, affordances, cognitive load, states, flow integrity) - Skip for backend-only features
1c. Generate Project Context (after first few phases)
-
/intel— populates.claude/context/with architecture, decisions, conventions, patterns - Skip for brand-new projects — run after there's enough code to analyze
2. Scaffold
- In the same session: "Let's scaffold the project using /scaffold"
- Confirm the scaffold plan (stacks, directory names)
- Claude creates subdirectories, git inits, bootstraps
- If Axon is installed, initial index runs automatically
3. Plan
- Start a new Claude session
- "Turn the design docs in docs/plans/ into a phased implementation plan"
- The planner reads the most structured artifact available: UX spec > PRD > brainstorm doc
- Review phases — small enough to review individually? Dependencies flow forward?
-
plan.mdsaved to project root
4. Execute + Review (repeat per phase)
-
cdinto sub-project (e.g.,~/prj/my-app/api/) - Open Claude Code session
-
/execute— picks up the plan, confirms the phase, implements on main - When done, start a new Claude session
-
/verify— checks the work against the plan, produces verification report, commits on PASS - Fix any FAILs, then move to next phase
-
/remember— capture any decisions, conventions, or patterns worth preserving
Or use hands-off execution:
-
/autopilot— for pre-planned phases: implement → test → verify → commit loop -
/orchestrator— for broader work: inline tasks, parallel execution, auto-phasing, model overrides - Both stop on FAIL — fix issues and resume with
/autopilot Nor/orchestrator N
5. Cross-Repo Handoff (if applicable)
- After API phases: "Generate an integration summary"
- Before web phases:
~/prj/polaris/context-pull.sh ../api - Then execute web phases with the integration context loaded
Ongoing Development
-
/recallat session start — loads relevant project context -
/wrapat session end — one-command close-out: memory, context, repo report (/rememberfor a single targeted save) -
/relaywhen context gets heavy mid-task — hand off to a fresh session instead of/compact
Single feature: branch → /execute → /verify → PR → merge
Parallel features: use git worktrees (the /worktrees skill)