AI Wizard

AI Wizard

AI Wizard produces an observed AI fluency profile. It is profile-first and score-second:

1. Profile and maturity stage
2. Evidence dossier
3. Primitive fluency map
4. Risks and missing evidence
5. Next-build plan
6. Secondary score

It is grounded in Vibe Pill methodology:

• Old AI Fluency Test dimensions: mental model accuracy, decomposition, failure recognition, tool-agnostic thinking, delegation judgment, iteration/refinement
• Vibe Pill primitives: system composition, pipeline thinking, tool thinking, integration thinking, orchestration/trust boundaries, feedback loops
• Meta-primitives: context engineering, state awareness, error recovery

Commands

python3 Skills/ai-wizard/scripts/ai_wizard.py scan
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --mode zo-native --depth capped
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --mode baseline --input Skills/ai-wizard/tests/fixtures/baseline
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --no-semantic
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --no-semantic --dogfood
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --semantic-provider zo --semantic-resume
python3 Skills/ai-wizard/scripts/ai_wizard.py profile --semantic-provider anthropic --semantic-resume
python3 Skills/ai-wizard/scripts/ai_wizard.py report --latest
python3 Skills/ai-wizard/scripts/ai_wizard.py history
python3 Skills/ai-wizard/scripts/ai_wizard.py submit --name "Your Name" --latest

Leaderboard

submit posts your latest profile score to the public leaderboard at https://va.zo.space/ai-wizard-rank — no keys or accounts required.

• The payload includes your three group scores; the server recomputes the public scoring formula and rejects totals that contradict their own groups. Consistent entries get a "consistent" badge; manual web-form entries without group scores are marked self-reported.
• Harness (zo / claude_code / codex / other) is auto-detected from the environment; override with --harness.
• --dry-run prints the exact payload without sending anything. --private submits without appearing on the public board. --run <dir> submits a specific run instead of the latest.
• The response reports your rank, percentile, and total entries. The board shows each name's best score.
• Override the endpoint with AI_WIZARD_LEADERBOARD_URL.

Defaults

• Local-first.
• Capped scan depth by default.
• Heuristic fallback by default; LLM adjudication is only active when --semantic-provider zo or --semantic-provider anthropic is requested and the profile reports semantic_status: complete.
• Semantic providers: zo uses /zo/ask (requires ZO_CLIENT_IDENTITY_TOKEN, Zo Computer only); anthropic uses the Anthropic Messages API (requires ANTHROPIC_API_KEY, works on any machine — Claude Code, Codex, plain shells). Model override: AI_WIZARD_ANTHROPIC_MODEL (default claude-sonnet-4-6).
• Quick scan is --no-semantic; balanced scan is default capped mode; deep scan is --depth full and may be paired with a semantic provider.
• Profile and report outputs stay local by default. Named semantic providers send redacted evidence packets only when explicitly selected; submit sends the documented score payload only when explicitly run.
• Public report redacts raw private evidence unless explicitly configured otherwise.
• Semantic provider failures, timeouts, and malformed responses are recorded in semantic_events; outputs still write with deterministic fallback scoring.
• --semantic-resume reuses semantic-reviews.jsonl checkpoints in the target run directory.
• --dogfood writes evaluator diagnostics (dogfood-report.json and dogfood-report.md) without raw private evidence.

Evidence sources (zo-native mode)

Three sources, allocated 60/30/10 of --artifact-limit and rebalanced when one under-delivers:

• Select sources with --sources workspace,conversations,git (default: all).
• Conversation traces are deduped, filtered of synthetic/system messages, and exclude ai-wizard's own sessions to prevent self-inflation.
• Environment overrides (all optional, colon-separated paths): AI_WIZARD_WORKSPACE, AI_WIZARD_ROOTS, AI_WIZARD_TRACE_DIRS, AI_WIZARD_ZO_MODEL, AI_WIZARD_ANTHROPIC_MODEL. Empty AI_WIZARD_TRACE_DIRS disables trace collection. On a machine without N5, roots fall back to scanning the workspace itself — no N5 layout required.

Outputs

Default outputs are written to Databases/ai-wizard/runs/<run-id>/, or to the supplied --out directory.

The local history database is Databases/ai-wizard/ai_wizard_history.sqlite.