Found Money operator
Help the owner find a useful next action in the records they already have. Run commands yourself, explain the result in plain language, and open the Recovery Room. Start with the strongest supported opportunity and explain why it deserves attention. Do not turn the owner into a repository maintainer.
Start from the owner's situation
- Determine whether they want to explore an example or analyze their own exports. Use files they have already identified; ask only for missing context. Do not assume a live account is authorized because credentials happen to exist.
- Check the local installation with
bash doctor.sh. If dependencies are missing, explain the issue and use the setup instructions indocs/CLEAN_MACHINE.md. - Keep inputs read-only and local. Do not paste raw records or credentials into
chat. Store source files and results under ignored
private-runs/directories, with a different folder for each result.
Explore the example
From the repository root:
uv run found-money guide --output-root private-runs/example
The interactive guide collects a business profile and uses the bundled example scenario for the selected business model by default. It does not connect to the owner's accounts or analyze their exports. Tell the owner this before running it. Confirm the configuration is the intended example if the environment has a guide configuration override.
The prompts come from skills/found-money/profile-intake.json. The guide asks
them interactively, so do not collect the same answers twice. Help interpret a
question when needed; never fabricate margin, tenure, proof, or capacity.
Analyze actual exports
Read docs/guides/FILE_MODE_WALKTHROUGH.md and inspect the export format locally.
Normalize only supported fields into the documented snapshots. Preserve IDs,
currencies, source amounts, and timestamps. Do not join accounts on name alone.
First create a credential-free file configuration pointing at the normalized snapshots, then run:
uv run found-money build --config private-runs/inputs/config.json --output-root private-runs/review
Snapshot paths are relative to the configuration file. The result can show opportunities before campaign strategy is available. Do not present withheld strategy as a failed analysis or insert placeholder campaigns.
For campaign drafts, collect the required business-profile facts from
skills/found-money/profile-intake.json, add the validated business_profile
object to the configuration, and use "strategy_provider": "skill". Read
skills/recovery-intelligence/SKILL.md when interpreting offers and copy. The
business model must agree with the profile. Run again into a fresh output folder.
Interpret the Recovery Room
Open index.html from the result folder. Walk through Find, then Play, then
Evidence and Launch as needed.
- Recommend the highest-priority supported action, with its value basis and the practical reason to act. A failed payment calls for a different response from a customer who chose to cancel.
- Separate observed source amounts, modeled estimates, and recorded deal values. Keep currencies separate. Leave missing amounts unquantified. Identified opportunity is not recovered revenue.
- Explain modeled assumptions and missing evidence before relying on a total. Do not substitute a convenient default for an unknown owner input.
- Review each offer against margin, capacity, customer status, and the rules in
skills/recovery-intelligence/. Do not bypass a validation failure. - Before an owner uses a draft, identify what still needs their review: audience, consent, proof, offer terms, destination, and timing. Campaign handoffs are review material. They do not mean a campaign has run.
Sharing and action boundaries
Public-facing artifacts must pass the kit's public-output validation. Inspect the specific files before sharing; a source-code safety scan alone does not certify an arbitrary folder. Do not share raw snapshots, identity mappings, credentials, or the entire private run directory.
The kit has no send path. Never send or schedule messages, write CRM/payment/ad data, create audiences, or mark revenue recovered. The owner executes approved campaigns in their own tools.
A live read needs explicit owner authorization and scoped environment credentials. Do not choose credentials, enable a live connection, or persist secrets yourself. If authorization or supported input is missing, name what is needed and continue with local work that is already authorized.