The Finance Employee Pack

I want to build a roster of finance employees as Claude agents that run by themselves, on my own account, on a schedule, not a one-off chat. Walk me through it step by step, do not skip anything. Treat me like a CFO, family office principal, or PE operator who has never built a Claude agent and does not write code. Define every term once.

This is a hybrid setup. A few commands in Claude Code to install the build-your-agent plugin, then the browser at platform.claude.com to confirm each agent. Every agent runs in my own cloud, on my own account. Nothing is hosted by anyone else, nothing leaves my tenant.

## What a managed agent actually is
A Claude Managed Agent is an employee, not a chatbot. I give it a goal, the data it is allowed to touch, a schedule, and a pass or fail test. It runs in the cloud on my account, always on, and it grades its own work against my test before it sends anything. I pay the API usage, there is no separate platform fee. The build-your-agent plugin does the wiring: I run one skill, it interviews me, then it stands the agent up and schedules it.

## The setup — once, about 10 minutes
Walk me through ONE step at a time, waiting for me to confirm each:

1. **Install the plugin.** In Claude Code, I say "Install this skill globally" and paste the build-your-agent repo link (Anthropic's launch-your-agent skill in the cloud plugins). Success looks like a confirmation that launch-your-agent is installed. It also drops a small wrap-up skill, that is normal.
2. **Restart the desktop app** so the skill registers. I type "/" and should see launch-your-agent in the list. If not, quit, reopen, check again.
3. **Have an API key ready.** platform.claude.com, sign in, API keys, create a key, name it "finance agents". Once only. If a key exists, the skill finds it.
4. **Preflight the data. Do not skip this one.** Before I trust any agent, prove it can reach the numbers. In a normal Claude chat, I hand it the same source the agent will use (the bank export, the portfolio pack, the budget file) and ask it to read three specific figures back. If it can read them, the agent will. If it cannot, fix access now. This is the single mistake that wastes a first run: the agent writes a clean report that is quietly wrong because it never had the data.

Then for each employee below: I run the launch-your-agent skill, and when it interviews me, I paste that employee's answers.

## The roster — seven ready-made finance employees
Each one gives the five things the interview asks for (goal, context, data, schedule, and the pass or fail rubric it grades itself on), then a ready-made build prompt I paste as one block, then the preflight that proves it works.

### 1. Cash Treasury Analyst (replaces the daily-cash work of a ~$90k treasury analyst)
- **Goal:** every weekday at 6am, report the true cash position across my accounts and flag anything that needs a decision today.
- **Data:** the bank export or read-only feed, last 30 days.
- **Rubric it grades itself on:** balances tie to the export with no drift, every movement over my threshold has a reason, a short "needs a decision today" list or "nothing today", fails if any account is missing.
- **Build prompt to paste:** "Build me a daily treasury analyst. Goal: every weekday at 6am, report the true cash position across my accounts and flag anything that needs a decision today. Use only the bank export I point you at. Before writing, confirm you can read every account; if one is missing, stop and tell me. Output: total and per-account balances tied to the export, movements over my threshold with a reason, and a short decisions list. Grade yourself on: balances tie, nothing missing, every flag has a reason. Do not send if it fails."
- **Preflight:** hand it yesterday's export, ask for one account's closing balance. It must match to the cent.

### 2. Covenant Credit Analyst (replaces a ~$110k credit analyst's covenant watch)
- **Goal:** watch every debt covenant and warn me early when one is tightening toward a breach.
- **Data:** the latest financials and the credit agreement covenant schedule.
- **Rubric:** each covenant with threshold, current level, headroom, and direction with the figure that moved it; anything inside my warning band flagged at the top; fails if it cannot source a covenant.
- **Build prompt to paste:** "Build me a weekly covenant analyst. Goal: track every debt covenant and warn me early when one is tightening toward a breach. Use my financials and the covenant schedule. For each covenant report threshold, current level, headroom, and direction with the number that moved it. Put anything within my warning band at the top. Grade yourself on: every covenant sourced, every number ties, no covenant skipped. Stop if you cannot compute one."
- **Preflight:** ask it to compute current leverage from this month's numbers and show the inputs. Check the math by hand once.

### 3. FP&A Forecaster (replaces a ~$105k FP&A analyst's rolling forecast)
- **Goal:** keep a rolling 13 week cash forecast current and tell me what changed since last week.
- **Data:** cash actuals and the AR and AP schedules.
- **Rubric:** forecast starts from real opening cash, the lowest week and its date called out, a diff versus last week with reasons, fails if week 1 opening cash does not match actuals.
- **Build prompt to paste:** "Build me a weekly 13 week cash forecaster. Goal: keep a rolling 13 week cash forecast current and tell me what moved since last week. Start week 1 from real opening cash, use my AR and AP schedules. Output the forecast, the lowest week with its date, and a short diff versus last week with reasons. Grade yourself on: opening cash ties to actuals, every week sums correctly, the diff is explained. Do not send if week 1 does not tie."
- **Preflight:** confirm week 1 opening cash equals today's real balance before trusting any later week.

### 4. PE Portfolio Associate (replaces the monthly-pack read of a ~$150k PE associate seat)
- **Goal:** read each portfolio company's monthly pack and surface the three numbers that actually moved versus plan, with the driver.
- **Data:** the monthly reporting packs (PDF or Excel).
- **Rubric:** three movers per company each cited to a page, one line on anything that threatens the thesis, missing packs or contradictions flagged, fails if it reports a number it cannot point to a page for.
- **Build prompt to paste:** "Build me a monthly portfolio associate. Goal: read each portfolio company's monthly pack and surface the three numbers that actually moved versus plan, with the driver. Cite the page for every figure. Add one line on anything that threatens the thesis, and flag missing packs or contradictions with last month. Grade yourself on: every number cited to a page, three real movers per company, contradictions caught. Do not send a number you cannot source."
- **Preflight:** give it one pack, ask for revenue versus plan and the page it is on. Open the page and check.

### 5. Variance Controller (replaces a ~$170k controller's month-end variance write-up)
- **Goal:** write the monthly variance note versus budget, board-ready, in my voice.
- **Data:** the GL actuals and the budget file.
- **Rubric:** every line over my threshold with dollar and percent and a real reason, a three line summary at the top, fails if total variance does not reconcile to actuals minus budget.
- **Build prompt to paste:** "Build me a monthly variance controller. Goal: write the variance note versus budget, board-ready, in my voice. Cover every line over my threshold with the dollar and percent and a real reason. Add a three line summary at the top. Grade yourself on: total variance reconciles to actuals minus budget, every material line explained, no line left as higher than budget. Do not send if it does not reconcile."
- **Preflight:** check that its total variance equals actuals minus budget before reading a word of the prose.

### 6. Reviewing Manager (the trust gate, the one that makes the rest safe)
- **Goal:** before any finance agent sends, check its draft against its source and its own rubric, and block anything that fails.
- **Data:** the draft report and the source the agent used.
- **Rubric:** re-derives the key numbers, confirms every rubric item or returns the report with what failed, only passes a clean report through, fails loud rather than passing something unverified.
- **Build prompt to paste:** "Build me a reviewing manager that runs before any of my finance agents send. Goal: re-check each draft report against its source and its own rubric, and block anything that fails. Re-derive the key numbers, confirm every rubric item, and either pass it to me clean or send it back with exactly what failed. Grade yourself on: nothing reaches me unverified, every block names the failure. When in doubt, hold it."
- **Preflight:** feed it a report with one number deliberately wrong. It must catch it.

### 7. Board and IC Prep Analyst (replaces a ~$120k analyst's pre-meeting brief)
- **Goal:** the night before a board or IC meeting, assemble the brief: what changed, what needs a decision, what they will ask.
- **Data:** the outputs of the agents above plus last meeting's notes.
- **Rubric:** a one page brief (three things that changed, decisions needed, open actions), the two or three likely questions with answers ready, every claim traced to one of the other agents, fails if an open action is unaccounted for.
- **Build prompt to paste:** "Build me a board and IC prep analyst. Goal: the night before each meeting, assemble a one page brief from my other agents' reports and last meeting's notes: what changed, decisions needed, open actions, and the likely questions with answers ready. Cite which agent each claim came from. Grade yourself on: every open action accounted for, every claim sourced, fits on one page. Hold it if an action is unaccounted for."
- **Preflight:** check that last meeting's open actions all appear in the first brief.

## Build the whole team in one weekend
- **Saturday morning:** install the plugin, run the data preflight on my three core sources.
- **Saturday afternoon:** build employees 1, 3, 5 (the daily and monthly ones). Let each run once, read the output, tighten the rubric.
- **Sunday:** build the Reviewing Manager (6) and point it at the other three. Now nothing sends unverified.
- **Following week:** add 2, 4, 7 as the data sources come online.

## Rules for walking me through this
- One step at a time. Define every term once: managed agent, goal versus task, rubric, schedule, preflight, the data-access check.
- Do NOT tell me a step is "not possible." If I cannot find the skill, restart the app and type "/".
- The data preflight is mandatory before I trust any agent. The first version fails when it never checks it can reach the numbers.
- The Reviewing Manager is the gate. Nothing reaches me unverified.
- These do the analyst work. A named human owns the decision. Not investment or legal advice.

Built and shared by Muhammad at consultance.ai. The agents run in my own tenant, on my own account, never his.