setup · private assay · no broad scan

Do not send a bundle until the scope is agreed.

The private beta uses manual bundles before the local command is ready. That makes the redaction boundary part of the product. The rule is simple: include only what explains Claude Code behavior, and exclude anything that authenticates, identifies third parties, or does not shape the agent.

01before export

Start with a file list, not a zip.

In the first exchange, send the shape of your Claude Code environment and the project you want assayed. The exact files come after that. This prevents accidental over-sharing and keeps the report tied to a real use case. 

# useful first message, no file contents yet
project: private app, TypeScript, 8 months old
Claude Code usage: daily
customized: CLAUDE.md, 4 agents, 3 commands, 2 hooks, 1 MCP
assay target: Claude over-explores and misses my intended finish standard
possible scope: current project .claude + selected ~/.claude command files

02include

Include files that shape Claude's behavior.

Usually useful

  • Project CLAUDE.md and local instruction files.
  • Project .claude/agents/, .claude/commands/, hooks, and settings.
  • Selected ~/.claude agents, commands, skills, hooks, and MCP configs.
  • Short notes about what Claude did wrong or needed repeated steering on.
  • One representative task transcript if you can redact it safely.

Sometimes useful

  • Repo tree, without file contents, when navigation is the issue.
  • Package scripts if commands change how Claude tests or builds.
  • Plugin manifests if plugins affect available tools or prompts.
  • Agent marketplace lists only when they affect your installed stack.

03exclude· no exceptions in beta

Never include secrets or broad personal archives.

Do not send

  • .env, API keys, OAuth tokens, SSH keys, private keys, credentials.
  • Cloud provider configs that contain account identifiers or secrets.
  • Customer data, employee data, private client names, or proprietary source files unrelated to Claude behavior.
  • Browser history, email, Slack, calendar, documents folder, downloads folder, photos.
  • node_modules, build folders, logs, caches, vendored dependencies.

Redact aggressively

  • Replace client names with [client].
  • Replace internal project names with [project] if they do not matter.
  • Replace file paths that reveal personal details with stable aliases.
  • Replace any credential-like value with [redacted], even if expired.
  • If you hesitate, remove it and say what category was removed.

04manual bundle shape

The current bundle is a folder of text files plus one manifest.

raising-agents-bundle/
  manifest.md
  project/
    CLAUDE.md
    .claude/
      agents/
      commands/
      hooks/
      settings.redacted.json
  user-claude/
    agents/
    commands/
    skills/
    settings.redacted.json
  notes/
    friction.md
    representative-task.redacted.md

The local command will eventually assemble this safely. The private beta keeps it manual so the product learns which files actually explain outcomes.

05manifest template

Every bundle needs a manifest.

# manifest.md

Assay target:
- What behavior should the report explain?

Included paths:
- project/CLAUDE.md
- project/.claude/agents/*
- user-claude/commands/*

Explicit exclusions:
- .env files
- keys/tokens
- customer data
- unrelated source files

Known redactions:
- [client] replaces customer name
- [project] replaces internal project name
- [path-home] replaces local home directory

Permission:
- You may use this bundle only to produce my Raising Agents assay report.
- Do not reuse it as training data.
- Delete local copies after report delivery unless we agree otherwise.

06request first

The correct next move is still the intake form.

Do the request first. The reply will tell you whether paste-only, current-project, or redacted-bundle mode is appropriate.

Request private assay See sample report