# AI Wizard

> Generate an observed AI fluency profile from real AI usage traces, workspace artifacts, and Vibe Pill methodology. Use when assessing how someone actually works with AI across systems, not when running a self-report quiz.

- Skill: `thevibethinker/ai-wizard` (Agent Skill, multi-file: 13 files)
- Install (CLI): `npx skillmds@latest add thevibethinker/ai-wizard`
- Raw SKILL.md: https://api.skillmd.com/api/skills/thevibethinker/ai-wizard/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Docs & Writing
- Author: thevibethinker (https://skillmd.com/u/thevibethinker)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/thevibethinker/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

```bash
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:

| Source | What it collects | Where |
|---|---|---|
| `workspace` | Build plans, skills, prompts (md/py/json/yaml/ts) | Auto-detected roots (`N5/builds`, `Skills`, `Prompts`) or `AI_WIZARD_ROOTS` |
| `conversations` | Real operator messages from AI session traces | `~/.claude/projects` (Claude Code JSONL), `~/.codex/sessions` (Codex rollout JSONL), `N5/logs/threads` exports, or `AI_WIZARD_TRACE_DIRS` / `--trace-dir` |
| `git` | Commit history (subjects + bodies) | Workspace git repo |

- 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`.