[01]Β LOOP
DocumentsintervalHardened

Client Docs to Forms

Inspired by Pollen β€” client documents in, completed forms out. AI ingests uploads from portal, email, or file drops, extracts fields with confidence scores, merges into one reviewed master record, routes judgment calls for human sign-off, and fills firm-specific output forms with a full audit trail.

Inspired by Pollen

by Trooper

Kickoff prompt
/loop 30m Start the "Client Docs to Forms" loop.

Inspired by Pollen (https://www.pollen.cx).

Goal: every open case has an approved master record and completed output forms ready for delivery or explicit escalation
Max iterations: 20
Between iterations run: Report open cases, pending extractions, low-confidence fields, master records awaiting sign-off, and unfilled output forms
Exit when: zero open cases without approved master record or explicit escalation, all low-confidence fields reviewed, and pending forms have owner or human approval

## Before you start
- Connect Gmail (required) β€” Client document intake from email
- Connect Notion (optional) β€” Master record and audit trail storage
- Attach Loop runner (required) β€” Self-pace iterations and run checks between passes.
- Attach Human approvals (required) β€” Queue external sends and high-impact changes for sign-off.
- Attach Customer operations (required) β€” Triage intake, onboarding, and support queues.
- Client upload portal (read via browser) β€” Secure document intake
- Firm output forms (write via manual) β€” Populate PDF/forms after sign-off

## Steps
1. Intake documents: Collect client documents from portal, email, or uploads; track new cases and attach source files. [tools: Gmail, Client upload portal]
2. Extract fields: AI reads documents with confidence scores per field; flag low-confidence extractions for review. [tools: Notion]
3. Build master record: Merge extracted fields into one reviewed record; resolve conflicts across source documents. [tools: Notion]
4. Human review: Route cases needing judgment to team sign-off before anything is final. [tools: Slack]
   Command: node scripts/doc-intake-report.js
5. Fill output forms: Populate firm-specific forms from approved master record; audit trail on every field. [tools: Firm output forms]

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.

Guardrails:
- Nothing is final without human sign-off on the master record
- Treat SSN, wages, EIN, and client PII as sensitive β€” minimize exposure in logs
- Do not auto-submit forms to external parties without approval
- Escalate fields below confidence threshold for human review
- Never use client data for model training; encrypted handling only
- Full audit trail on every field change

Paste the kickoff prompt into Cursor, Claude Code, or Codex. Deeplinks do not install hook files.

Steps

1. Intake documents

Collect client documents from portal, email, or uploads; track new cases and attach source files.

Tools: Gmail, Client upload portal

2. Extract fields

AI reads documents with confidence scores per field; flag low-confidence extractions for review.

Tools: Notion

3. Build master record

Merge extracted fields into one reviewed record; resolve conflicts across source documents.

Tools: Notion

4. Human review

Route cases needing judgment to team sign-off before anything is final.

Tools: Slack

node scripts/doc-intake-report.js

5. Fill output forms

Populate firm-specific forms from approved master record; audit trail on every field.

Tools: Firm output forms

Flow diagram

Guardrails

Rules the agent must follow so it cannot cheat the exit condition.

  • Nothing is final without human sign-off on the master record
  • Treat SSN, wages, EIN, and client PII as sensitive β€” minimize exposure in logs
  • Do not auto-submit forms to external parties without approval
  • Escalate fields below confidence threshold for human review
  • Never use client data for model training; encrypted handling only
  • Full audit trail on every field change

More Documents loops