Generate and cluster synthetic opinions from hundreds of local AI personas — a CLI for testing messages before you publish.
Status: alpha (0.x). It works end to end and is tested, but commands, output and artifact schemas may still change between minor versions.
Give it a draft (an ad, a post, an announcement). Wind Tunnel shows it to a demographically grounded persona pool, generates their reactions with a local LLM, and maps the crowd Pol.is-style — opinion groups, consensus, division and the minority view — fully local via Ollama, nothing leaves your machine.
$ wt-cli run "Introducing a 4-day work week. However, salaries will be reduced by 10%."
Backlash index ████████████████████░░░░ 85 / 100 HIGH
Voices ████████████████████████ critical 83% (5) · neutral 17% · favorable 0%
Triggers
1. "salaries will be reduced by 10%" (High) → employees in structured,
mission-critical, or low-margin industries
◆ Pay-First Skeptics (4) ████████████████
They believe that reducing pay to shorten the work week is inherently
disrespectful and fails to address the real issues workers face.
「I don't like it at all—cutting pay just to work fewer days is plain
foolish. You can't expect people to do more with less…」
— 79 · not_in_workforce · Lincoln, AR
...- Backlash index & voice split — how hot it runs, at a glance
- Triggers — which wording offends which segment
- Opinion groups — the camps that form, each with its belief and real member voices, plus a minority report on what the majority overlooks (a unanimous crowd is honestly shown as one camp)
- Rewrites — alternatives that keep your intent but defuse the heat
wt-cli detail— every voice in full and the proposition × group table- Plain artifacts per run (JSONL / JSON / CSV) for pandas, R, or Excel
- Sample — N personas are drawn from the demographic pool (country / region / age / sex filters).
- React — each persona answers in character (fast bulk model, parallel batches), framed by the situation: an anonymous board runs hotter than real-name social media.
- Score — every reaction is classified as stance × unsigned intensity and the signed score is composed in code, so a small model can never mis-sign a criticism as praise.
- Cluster — Pol.is-style opinion-space analysis with LLM-inferred votes: propositions are extracted from the corpus, every reaction votes agree/disagree on each, and the vote matrix goes through PCA + k-means. Consensus, divisive and minority views fall out of the group × proposition agreement rates — and a silhouette check collapses to one camp rather than inventing structure in a unanimous crowd.
- Verdict & rewrites — the analysis model reads the aggregate stats plus a stratified sample weighted toward both extremes for the backlash index and triggers; rewrites target those triggers specifically.
Every JSON hop is constrained-decoded (schemas shipped as Ollama's format),
so an 8B model cannot break the pipeline with malformed output. The full
reasoning lives in docs/DESIGN.md.
Requires Node >= 20 and Ollama:
brew install ollama && brew services start ollama # macOS (or the desktop app)
ollama pull qwen3:8b && ollama pull qwen3:14b # role models (one-time, ~15 GB)
npx wind-tunnel-cli doctor # verifies the setup, tells you exactly what's missing
npx wind-tunnel-cli personas pull usa # streams a persona pool from Hugging Face (~30 s)
npx wind-tunnel-cli run "draft copy..."
Installed globally (npm i -g wind-tunnel-cli) the command is wt-cli. Runs are
resumable (wt-cli resume <run-id>) and every option — countries, audience
filters, situations, model overrides — is in the
command reference.
Persona pools cover 8 countries (USA, Japan, India, Brazil, France, Korea,
Vietnam, Belgium), built on NVIDIA's
Nemotron-Personas
datasets (CC BY 4.0); custom pools plug in via --personas-file.
| Commands & configuration | every command and option, config.toml, env vars, run artifacts |
| Docker | Linux + NVIDIA compose setup, image usage |
| Design | architecture and the decisions behind it |
| Testing | the four-layer test strategy, coverage, what deliberately isn't automated |
| Releasing | maintainer release checklist |
pnpm install && pnpm test && pnpm build
node packages/cli/dist/index.js doctor # or: cd packages/cli && npm link
Ollama-dependent tests skip themselves without a daemon; building first also enables the black-box E2E of the binary. Strategy: docs/testing.md.
Apache-2.0. Bundled third-party licenses: THIRD_PARTY_LICENSES.md. Persona datasets are CC BY 4.0 and fetched by users directly from Hugging Face, not redistributed here.