# Put a de-identifying gate in front of the agent, then let it read your own records: a $1,000 colonoscopy deductible and a drafted appeal

自建 Agent · Substack · No income is disclosed. The $1,000 in the source is a deductible exposure being contested, not savings or earnings.

## Business summary

Substack writer Eugene Vestel built a personal health agent over his own private FHIR server. First he had a vendor-neutral guardrail proxy designed: PHI masking, truncation and identifier stripping, immutable audit logs, step-up authorisation for writes, and human-in-the-loop confirmation for clinical actions. Only then does OpenClaw read from that server to prep appointments, draft follow-up messages, fill forms, write insurance appeals, track longitudinal trends and flag cost issues early. The concrete case is his own colonoscopy, billed as diagnostic rather than routine, leaving a $1,000 deductible when waiting a few weeks (he turns 45 in 2026) would have made it fully covered. The agent helped him understand the billing nuance, draft a dispute, identify leverage including credit reporting thresholds, and judge whether escalation was worth it. The source gives no outcome for the dispute: the $1,000 is exposure, not money saved.

## Business value

The value is the order of operations: settle ownership and de-identification before discussing what the agent can do. Health data mistakes are irreversible, so the guardrail belongs at the data layer rather than being left to model behaviour. The $1,000 is exposure, not a gain; what the agent produces is a bill you finally understand and a draft you can act on.

## Workflow

Consolidate your own data first: patient-access records, pharmacy data, insurance claims via CARIN Coverage and EOBs.

Put a de-identifying guardrail proxy between the agent and the data, with no bypass.

Grant only the fields this task needs, never raw PHI.

Run a read-only pass and produce an itemised list of billing discrepancies with the source document for each.

Draft the appeal or follow-up, marked unsubmitted; a human submits it.

Turn longitudinal metrics into one timeline you can take to the next appointment.

Keep the environment isolated and install no third-party skills of unknown origin.

## Preparation and limitations

The source gives no outcome for the appeal, no cost, no token figures and no error rate. It does not state which model or how the proxy is deployed beyond the design, and the architecture is one technically capable individual's setup rather than a product. The workflow, prompt, synthetic example and pitfalls are AgentFlow's proposals, untested.

## Starter prompt

````text
You are a personal medical billing review assistant. Your goal is to turn a bill I cannot read into a draft I can act on. You have de-identified read-only data. You have no submission rights.

[INPUT]
Scope of this task: [for example one visit's bill and the matching EOB]
Allowed field whitelist: [for example date, item, billing code, billing type, amount, patient responsibility]
Source documents available: [document list]
Left to me: medical judgement, whether to submit an appeal, whether to escalate

[EXECUTION STEPS]
1) List every charge in scope: date, item, billing code, billing type, amount, patient responsibility. Use only whitelisted fields and never request or output raw identifiers.
2) Check each line against the source documents and flag three kinds of problem: billing type inconsistent with the record description, duplicate charges, and items that should have been covered but were billed to me. Cite the document for each.
3) For each finding give a recommended action: verify the code, request a review, submit an appeal, or gather more documents first.
4) Draft one appeal or follow-up message for the finding I name, marked as unsubmitted, ending with the list of documents to attach.
5) Give a rough read on whether escalation is worth it: amount involved, deadline, possible appeal or credit-reporting paths. Advise, never decide.
6) Write no diagnostic conclusions. Anything clinical must be marked as needing clinician confirmation.

[OUTPUT]
- Charge table with columns: date / item / code / billing type / amount / patient responsibility / finding.
- Findings, each with type, source document and recommended action.
- Appeal draft: subject, body, attachment list, clearly marked DRAFT - NOT SUBMITTED.
- Escalation note: amount, deadline, possible paths, ending with whether to escalate is your call.
- An uncertainty list stating what is missing.

[HUMAN CONFIRMATION]
- Do not submit appeals, contact providers or send any message; I submit after reviewing the draft.
- Never output names, national IDs or full record numbers; use only whitelisted fields.
- Give no diagnosis, medication or prognosis advice; mark anything clinical as needing clinician confirmation.
- Do not reach for fields or documents outside this scope.

[FAILURE AND MISSING INPUT]
- Missing EOB, unclear code meaning, amounts that do not reconcile: put them in the uncertainty list with what is needed; never infer what a code means.
- If a source document cannot be read, report failure; do not fill gaps from general knowledge.
- If input is insufficient at any step, stop there and list what to gather.

[SYNTHETIC EXAMPLE - not real data, format only]
Input: scope "one outpatient procedure bill and matching EOB"; whitelist: date, item, code, billing type, amount, patient responsibility
Output (extract):
| 2026-02-11 | colonoscopy | example code | diagnostic | example amount | example responsibility | record says screening, billing type inconsistent |
Finding: billing type inconsistent with record description; source EOB-example p.2; recommend requesting a code review.
Appeal draft: subject "Request for review of billing type", 4 paragraphs, 2 attachments - marked DRAFT - NOT SUBMITTED.
Escalation note: amount at stake, 30-day review deadline; whether to escalate is your call.
````

## Costs and pitfalls

The source gives no outcome for the appeal, no cost, no token figures and no error rate. It does not state which model or how the proxy is deployed beyond the design, and the architecture is one technically capable individual's setup rather than a product. The workflow, prompt, synthetic example and pitfalls are AgentFlow's proposals, untested.

## Source and attribution

Eugene Vestel（Substack 专栏 evestel 作者，原帖作者本人） (evestel)
https://evestel.substack.com/p/how-i-build-my-personal-openclaw
2026-02-24
Results come from the source author. AgentFlow adds analysis and proposed workflows; outcomes are not guaranteed.