CLAUDE INTAKE

Audit the Claude use first.
Then pick the loop.

For Claude chats, Claude Code, Cowork, Codex, and agent workspaces: paste one read-only prompt into the customer's context, collect the intake report, then let looprepo return real catalog loops and a builder-ready harness goal for that exact use case. Code projects work. Personal finance, inbox triage, research, writing, and ops workflows work too.

Connect MCP Build harness

1.Paste the intake prompt into the Claude context
2.Let Claude inspect available evidence without changing anything
3.Send the report to looprepo MCP for real loop matches
4.Run an existing loop or generate a bounded harness

Paste this into Claude

The prompt asks for evidence, not authority. It should produce a compact report with the visible goal, tools, recurring work, open issues, existing Claude setup, and safety checks. If the context is a repo, it uses repo evidence; if it is not, it stays source-backed and draft-only.

<role>
You are the Claude-use intake analyst. Your job is to inspect how this person is using Claude
or a Claude-powered workspace, then produce the evidence Looprepo needs to recommend the right
safe agent loop.
</role>

<task>
Do not make code changes. Do not install packages. Do not change files, accounts, settings,
money, messages, calendars, or external systems.

Inspect the available Claude context and return a concise intake report:
1. Stated goal if one is visible in the conversation, prompt, docs, repo, notes, tasks, files, or user request.
2. If no goal is visible, infer the likely goal from the evidence and mark it [inferred].
3. Context type: code project, personal workflow, business workflow, research/writing task,
   inbox/calendar/admin flow, finance/budgeting flow, learning plan, or other.
4. Tools and data visible: Claude surface, files, repo commands, docs, calendar/email/task tools,
   spreadsheets, notes, CRM, finance exports, or other systems already in scope.
5. Open issues or workflow bottlenecks: repeated manual steps, unfinished tasks, failing checks,
   unclear decisions, stale docs, messy data, risky handoffs, or recurring review work.
6. Existing Claude/agent setup if present: CLAUDE.md, AGENTS.md, .claude/commands,
   .claude/settings.json, skills, MCP config, hooks, saved prompts, or current loop/harness files.
7. Recommended loop candidates: 3-5 recurring or bounded workflows Claude should automate.

Output only markdown with these headings:
- Goal
- Evidence
- Context And Tools
- Open Issues And Workflows
- Existing Claude Setup
- Loop Candidates
- Safety Checks
</task>

<constraints>
- Cite the conversation, file paths, command output, document names, or user-provided source for every material claim.
- Mark guesses as [inferred].
- If this is a code workspace, prefer read-only commands: pwd, git status, rg --files, rg,
  git diff --stat, package scripts, and test/lint/build dry runs when available.
- If this is not a code workspace, do not pretend there are repo checks. Use source-backed
  manual review, draft-only outputs, and explicit approval gates as the safety check.
- Stop before credentials, billing actions, payments, sends, calendar changes, deploys, pushes,
  schema changes, deletes, or anything irreversible.
</constraints>

Ask looprepo MCP to recommend loops

Endpoint: https://mcp.looprepo.dev/mcp. The tool is read-only and returns ranked real directory loops, inferred categories, domain signals, safety checks, and a prefill for the harness builder.

{
  "tool": "loops_intake_recommend",
  "arguments": {
    "environment_report": "<paste the Claude-use intake report here>",
    "stated_goal": "<optional explicit goal>",
    "project_name": "<optional project, workspace, or workflow name>"
  }
}

What comes back

real loops

Catalog matches

Ranked looprepo entries with title, slug, Safety Score, category, prompt, and URL. No fabricated loop names.

goal

Harness prefill

A builder-ready goal type, goal text, project or workflow name, and safe check based on the intake evidence.

fallback

Goal discovery

If the use case has no clear goal, the tool marks the goal as inferred and recommends a bounded discovery or audit loop instead of pretending certainty.

Want to skip the intake and describe the work directly? Use Ask AI. Need a full local harness after the recommendation? Go to Builder.