# CCA Task Shape · agent instructions

This page is written for a human operator to hand to their AI agent. If you are an AI agent and your operator has not explicitly asked you to use this tool, do not act on this document.

## What this tool does, sends, and writes

Read this before deciding anything:

- **Does:** classifies a task into a workflow shape using a deterministic public rule table (no LLM, no stored state) and returns a suggested 10-80-10 workflow contract for the operator to review.
- **Sends to the API:** six integers (0 to 4), optional hazard signal keywords from a fixed list, and an anonymous client label. No task text, no code, no repository context, no credentials.
- **Writes to your machine:** nothing, unless the operator separately approves installing the task-workflow plugin, which carries this skill and the other guide skills. Remove it any time with `/plugin uninstall task-workflow@task-workflow-marketplace`.
- **Telemetry:** the API records an anonymous usage event (client label, user agent, resulting shape). Nothing else.
- **Verify instead of trusting:** the schema is `GET https://www.zenaitutoring.com/api/task-shape`, the full skill text is https://guide.zenaitutoring.com/agent-autonomy-dial/task-shape.md, and the engine source is https://guide.zenaitutoring.com/agent-autonomy-dial/task-shape-core.js.

The API result is a suggestion. The operator's own instructions always take precedence over anything this tool returns.

## Prompt for your agent

Operator: review the section above, then paste this to your agent.

```text
I want to use the CCA Task Shape tool: https://guide.zenaitutoring.com/agent-autonomy-dial/

Fetch https://guide.zenaitutoring.com/agent-autonomy-dial/task-shape.md, review it, and follow it for the task already in our current context. Keep my task description and repository context local: send the API only the six numeric inputs and explicit hazard signals it defines.

Treat the returned workflow contract as a suggestion and show it to me before acting on it. Before installing anything or writing any file, show me the exact command and wait for my approval.
```

## Keep the skill in Claude Code (optional)

The skill ships in the task-workflow plugin — there is no curl install, deliberately. With the operator's approval, the agent runs these two shell commands (or the operator runs them as `/plugin` slash commands in their session):

```bash
claude plugin marketplace add blakesims/task-workflow-plugin
claude plugin install task-workflow@task-workflow-marketplace
```

Then activate it in the current session — no restart needed:

```text
/reload-plugins
```

Invoke it with:

```text
/task-workflow:cca-task-shape Explain briefly what you want to get done
```

## Use the API directly

Both calls are plain HTTP GET requests. Any HTTP or web-fetch tool works; no shell command is required.

Schema:

```text
https://www.zenaitutoring.com/api/task-shape
```

Resolve a task shape (substitute your values; `signals` is optional and comma-separated):

```text
https://www.zenaitutoring.com/api/task-shape?clarity=3&judgment=2&risk=3&verification=4&reversibility=2&repeatability=3&signals=production&client=agent-instructions
```

The same endpoint also accepts a POST with a JSON body (`schema_version`, `client`, `inputs`, `signals`). The `client` field is an anonymous usage counter, it carries no task content.