Claude Code for Finance Systems

The course

About this course

For finance-systems people who script and have already built one governed agent. The morning takes one process from the team’s list all the way: a connection to the source instead of an upload, the deterministic core as code with tests, an MCP link to a system where it helps, an approval queue, a log, and a schedule. In the afternoon each person builds a second process alone from its card, with a partner reviewing. The last hour writes the team’s conventions: the rules file Claude Code reads, the repository layout, how credentials are handled, what gets committed, and the README standard that lets a colleague run any agent cold. The team leaves with two agents running and the standards to build the next ten. This course runs on Claude Code.

Who it is for

Finance systems, reporting, and data people who already script, and who have taken Build a Governed Agent or run something equivalent

Format

Full day for a small group. One process built end to end in the morning with the full apparatus; each person builds a second alone in the afternoon; the day ends with the team's conventions written down.

What the team leaves with

Outcomes

• One process running end to end: connected to its source, deterministic steps in tested code, AI for exceptions and narrative, an approval queue, a log, a schedule.
• A second process built by each person alone in the afternoon, reviewed by a partner, and running.
• The team’s Claude Code conventions in a repository: rules file, layout, credential handling, commit discipline, README standard.
• The judgement to scope the next ten: which of the team’s processes belong in code, which belong to the AI, and which stay with a person.

Outline · 6 hours

How the session runs.

Usually 9:30 to 12:00 or 2:00 to 4:30. Other times are fine.

1. 0130 min

   The boundary, the setup, the standards

   • One process from the team’s list. Inputs, outputs, who approves, where it logs.
   • Project setup and the rules file that makes Claude Code behave like a colleague who knows the platform.
   • What credentials never look like. A real example, sanitized, from an archive.
2. 0260 min

   Connection, code, tests

   • A query or a watched folder instead of a paste.
   • The deterministic core, sums, joins, as-of logic, thresholds, as a script with tests Claude Code writes and you review.
   • The AI layer on top: read the exceptions, classify, draft the note.
3. 0345 min

   Systems, approval, log, schedule

   • An MCP link to a system where it helps; a read-only boundary where it does not.
   • Output to a review queue, never a ledger. The approval is the trigger for anything that posts.
   • Every run writes what it did and why. Scheduling it, and the morning digest.
4. 0445 min

   Run it, break it, fix the card

   • First scheduled run in the room. A planted failure. The log finds it.
   • Fix the card and the rules, not the prompt. Rerun.
   • A second person runs it from the README.
5. 0590 min

   Your second process, alone

   • Each person takes a card from the team’s list and builds the agent solo.
   • Partners review each other’s code, log, and approval gate at the halfway mark.
   • Neil circulates. Nobody gets code written for them.
6. 0660 min

   The team’s conventions

   • The conventions repo, written together: rules file, layout, credentials, what gets committed, README standard.
   • Where the agents live, who owns each, how a new one gets added.
   • What to do when it gets bigger than a script.
7. 0730 min

   The next ten

   • The team’s list, re-ranked with what it now knows.
   • Which belong in code, which to the AI, which stay with a person.

Bring

What to bring

• A laptop with a terminal, Claude Code installed, and a Claude login your organization allows for work. Setup instructions come ahead of the course.
• Build a Governed Agent, or the equivalent: you have built at least one scheduled, logged agent with an approval gate.
• The team’s ranked list and sanitized inputs for two processes on it.