Process Radar

Try the demo → — Describe your task and find an AI process that fits. No setup required.

View the talk slides (PDF).

A local demo for choosing a way to organize agentic work. Describe a task, get an LLM's suggested process, and compare its seven-dimension radar with the talk's six reference profiles. The speaker cheat sheet explains every reference score.

The numbers are Dustin's intended-use characterizations. They are a discussion aid, not benchmarks or a current product capability audit. A larger radar is not a better process.

Use it as a skill

The portable Process Radar skill estimates nine task-context answers from available context, derives dimension scores from those answers, then uses the scores to recommend a process. It does not require extra user input before scoring. It optionally charts the assessment and can review explicitly selected past conversations for process mismatches. It runs inside Claude Code, Codex, or another compatible harness using that harness's own model. No separate API key or web server is needed.

See installation and examples. Conversation review is opt-in, scoped, and read-only; installing the skill does not start scanning history.

Run

Requires Node.js 22 or later. There are no package dependencies or build step.

npm start

Open http://127.0.0.1:4317. Start with the blank “I want to…” prompt. Recommendations need a connected model; the cheat sheet remains available under Setup without one.

The default configuration uses GPT-5.5 through OpenAI. For recommendations, copy .env.example to .env if needed, fill in OPENAI_API_KEY, and restart. The prepared local checkout already contains a key-free .env:

cp .env.example .env
# Edit .env with your preferred editor.
npm start

Setting	Meaning
LLM_BASE_URL	OpenAI-compatible base URL ending in /v1, such as http://127.0.0.1:11434/v1 or https://api.openai.com/v1. Use your server's actual address.
LLM_MODEL	Exact model ID offered by that endpoint. The example uses gpt-5.5, as requested.
OPENAI_API_KEY	Your OpenAI key. Only used for api.openai.com; never forwarded to another endpoint.
LLM_API_KEY	Override key for any configured provider. Empty is fine for an unauthenticated local server.
LLM_JSON_MODE	true sends response_format: {type: "json_object"}. Set false if your endpoint rejects it.
LLM_TOKEN_FIELD	max_tokens for many compatible local servers; max_completion_tokens for OpenAI reasoning models.
LLM_MAX_TOKENS	Output cap, default 4096. Reasoning models may need a larger cap because it includes reasoning.
LLM_REASONING_EFFORT	Optional. The GPT-5.5 example uses low and an 8192-token cap. Clear it for local servers without support.
LLM_TIMEOUT_MS	Request timeout, default 120000.
PORT	Local app port, default 4317.

The server uses POST /chat/completions. It expects a non-streaming choices[0].message.content JSON string. Local server compatibility varies. JSON mode ensures JSON syntax where supported; the app separately validates all seven IDs, score increments, explanations, and recommendation IDs. It reports malformed, incomplete, refused, or timed-out responses without manufacturing a fallback recommendation or automatically retrying a paid request.

The API format follows the OpenAI Chat Completions reference and structured-output guidance. Chat Completions is used here for broad endpoint compatibility.

The GPT-5.5 model documentation lists Chat Completions support. To switch to a local server, change the base URL and model, clear reasoning effort if unsupported, and use the token-limit field your server accepts. No code change is needed.

Use it in the talk

1. Connect a local or hosted model before the session.
2. Describe your task in the free-form box. Include any checks you trust, human involvement, uncertainty, or constraints that matter.
3. Click “Find my approach” to reveal the recommendation, practical first steps, and radar. The green shape is the suggested way of working; dashed blue is that approach's reference profile.
4. Expand “Reasoning & alternatives” for assumptions, questions, the tradeoff, and score explanations. Revise your description and rerun as needed.
5. Use “Save chart” for an SVG, or save the recommendation as JSON from the expanded reasoning.

The LLM chooses a contextual starting point. There is no total-score, polygon-area, or distance-based winner. “Needs Strong Verifier” measures dependence on automatic success checks; a task with weak automatic checks should usually get low dependence, even when it is high-stakes.

Edit the rubric

• data/catalog.json: one source for labels, definitions, baseline scores, and rationales.
• docs/cheat-sheet.md: generated speaker notes with all 42 score rationales.
• model.mjs: LLM prompt, endpoint adapter, and response validation.
• skills/process-radar/scripts/radar-svg.mjs: SVG renderer shared by the web demo and portable skill.
• skills/process-radar/: self-contained installable skill, chart helper, and history extractor.

npm run docs
npm test

The baseline retains Plain Harness. The proposed Skill-guided Coding rename and revised scores are recorded separately in pendingRevision; they do not participate in live recommendations. The source chart had acquired that new title while still carrying the earlier scores. Workflows' verifier score of 25 and Plain Harness' human-worker score of 50 remain explicitly flagged for review.

Privacy and limits

• The app binds only to 127.0.0.1. It is a single-user local demo, not a hardened public service.
• API keys stay in the server environment. .env is ignored by Git and never served. Do not paste keys into task descriptions.
• Task text is sent to your configured endpoint when you submit it. A local app can still use a remote model: check the endpoint before entering sensitive information.
• The app does not persist task descriptions, responses, or usage logs. Your model provider may have its own logging and retention policy.
• One recommendation runs at a time. Requests have input/output size caps, a timeout, and an output-token cap. These do not guarantee a financial spending limit.
• No external scripts, fonts, analytics, automatic model downloads, or account setup.
• The model can misread a task or rationalize a poor choice. The reference scores do not establish shipped product capabilities, improved model performance, or safe deployment for consequential work.

Tests use a deliberately fake OpenAI-compatible endpoint. Passing them verifies the integration and validation; it does not evaluate recommendation quality from any real model.