Raw: https://promptreports.ai/docs/agents.md

# How agents use the docs

An agent with no prior setup will go from a domain name to a finished run in eight steps.

1. **Find.** Fetch /llms.txt. It names the docs index, the use case index and the MCP server.
2. **Learn.** Read /docs/concepts.md once. Every other page uses its words.
3. **Choose.** Fetch /use-cases/index.json and pick a workflow; read its /use-cases/<id>/openspec.md.
4. **Connect.** Add the MCP server. Sign in through the browser with a Client ID Metadata Document.
5. **Check.** Call get_workflow to see which products the org has connected. Tell the person what is missing.
6. **Run.** Call start_run. Poll get_run, or wait for the run.waiting_approval webhook.
7. **Hand off.** When a step waits for approval, tell the person where to decide. Never try to approve it yourself.
8. **Report.** When the run ends, return the record link and the outcome as met, not met or not measured.

## What an agent session looks like

```
agent  GET https://promptreports.ai/llms.txt
agent  GET https://promptreports.ai/use-cases/trust-packet/openspec.md
agent  start_run({ workflow: "trust-packet", inputs: { buyer: "Contoso Health" } })
runner R-2045 running; step 5 waits for file.share.external
agent  "Dana, the trust packet for Contoso is ready to share. Approve it at
        app.promptreports.ai/approvals. Three sections are not built yet and
        show as missing."
person approves
agent  get_run({ run_id: "R-2045" }) -> done, record T-8G0B3N
```