Claude Financial Services: what the public plugins actually provide
Build a reviewed handoff inbox for Financial Services
A proposed learning project turns validated requests into reviewable records without giving documents control of routing.
What you will learn
- Make the first extension small
- Define the acceptance cases
- Connect only after the offline contract holds
Before you start
- Command-line and JSON familiarity
- A disposable workspace with synthetic financial records
A proposed learning project turns validated requests into reviewable records without giving documents control of routing.
Key takeaways
- This is a proposed extension.
- Deduplication and authorization are distinct.
- Offline success does not prove production suitability.
Make the first extension small
Build an offline inbox for synthetic handoff events. Store a task identifier, sender identity, target slug and event payload in separate fields. Display the requested action and its evidence without executing it.
The project extends the reference architecture conceptually. It is not an existing feature of the inspected repository, and passing the exercise does not qualify the result for handling real customer records.
Define the acceptance cases
Require complete nested JSON to parse, reject unknown targets and extra payload fields, and keep malformed or incomplete messages in a separate rejected queue. Use escaped braces inside a string to check that framing follows JSON structure.
Give every accepted event a deduplication key and a visible pending-review state. A repeated event must not create a second authorized action. Record why a reviewer accepts or rejects it, while keeping sensitive source data outside the event body.
Connect only after the offline contract holds
Run the parser and state-transition tests before adding a model or connector. Then review how the chosen transport authenticates senders and how a failed acknowledgment is retried. These are different questions from whether the JSON is valid.
Finish with a short demonstration using invented records: one accepted request, one duplicate and one rejected target. The deliverable is a testable boundary and an audit trail; live deployment, financial correctness and compliance need their own reviews.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Implement a passive inbox for synthetic events.
- 2
Test framing, targets, duplicate keys and review state.
- 3
Demonstrate acceptance and rejection without external actions.
Copy-ready example
{"task_id":"synthetic-001","sender":"fixture","target":"gl-reconciler","state":"pending-review","dedup_key":"synthetic-001:1"}Frequently asked questions
Must the exercise call a real model?
No. Deterministic synthetic events are enough to test the control boundary first.
What should the demonstration contain?
An accepted draft event, a duplicate and a rejected target, with explicit reviewer decisions.
Sources
- Financial Services / README.mdSource checked 2026-09-23
- Financial Services / scripts/orchestrate.pySource checked 2026-09-23
- Financial Services / scripts/validate.pySource checked 2026-09-23