arthai-marketplace

Tutorial: Calibrate a Project

In this tutorial you’ll run /calibrate on a project for the first time and verify everything it created. Calibration is how the toolkit stops being generic: it deep-learns your codebase, seeds a knowledge base, and configures agents, skills, and MCP servers to match how your project actually works.

Time: ~15 minutes (most of it is the calibration running) You need: the toolkit installed in a real project (Getting Started).

Step 1: Install the toolkit (skip if done)

If you haven’t installed yet, follow Getting Started Steps 1–6: activate your license key, add the arthai-marketplace marketplace, install a bundle (start with prime), and /reload-plugins.

Step 2: Run calibration

In Claude Code, inside your project:

/calibrate

First you’ll see a mode line explaining which mode was picked and why — with no existing profile, that’s a full calibration. Then the phases run:

  1. Scan — a scanner agent deep-reads the project: platform, architecture patterns, coding conventions, testing patterns, domain model, integrations, environments.
  2. Evaluate — an evaluator scores every toolkit agent, skill, and hook against the scan and identifies gaps.
  3. Profile — a profiler writes .claude/project-profile.md.
  4. Recommend — the report is presented and you’re asked: Install all / Pick items / Skip.

If your CLAUDE.md is missing or still has <!-- TODO --> placeholders, calibrate runs /scan first to populate it.

Step 3: Approve the recommendations

The Phase 4 prompt is a real checkpoint — nothing is installed without your approval. The report covers recommended MCP servers, toolkit categories, workflows, and custom agents/skills tailored to your stack. For a first run, Install all is the simplest choice; Pick items opens a multi-select if you want to be choosy; Skip writes the profile only.

After you approve, the install phase runs in parallel: the installer applies your approved items, a knowledge agent seeds .claude/knowledge/, and a best-practices agent writes .claude/knowledge/shared/best-practices.md. A final verification phase confirms everything landed and prints a completion report. An independent verifier then reports PASS / GAPS FOUND / INCONCLUSIVE — gaps don’t block; you decide whether to rerun.

Step 4: What got created

Artifact What it is
.claude/project-profile.md Architecture, conventions, domain model — the file agents read at runtime
.claude/knowledge/ The seeded knowledge base, including shared/best-practices.md
.claude/settings.json Recommended MCP servers added (never removed)
.claude/agents/, .claude/skills/ Any custom agents/skills created, as regular files you own
.claude/monitors/ Pre-adapted Monitor configs for your CI/deploy platform, with activation steps printed (Monitors)

Step 5: Start a fresh session

Calibration is the heaviest context operation in the toolkit — everything of value is in the files, so start a fresh Claude Code session when it finishes. At the start of that next session, the knowledge graph is built automatically from your knowledge base: a ranked index of your conventions, domain rules, patterns, and vocabulary that workflows like /fix, /planning, /implement, and /qa query to pull in only the most relevant context. The graph auto-rebuilds whenever your knowledge base changes — no action needed from you.

Step 6: Verify it worked

In your terminal:

ls .claude/project-profile.md
ls .claude/knowledge/shared/

You should see the profile and the seeded knowledge files. Then, in a Claude Code session:

/calibrate status

This prints the current calibration state without spawning any sub-agents. If the report ever says calibration_status: partial, a phase failed and its dependents were skipped — re-run /calibrate full.

Finally, run /onboard — with a profile in place you get a project briefing grounded in what calibration learned.

What you learned

Next tutorial: Plan and implement a feature — the full PRD → plan → build → PR workflow.