AI Engagement Team
Five Claude skills that take a small firm from a first call to working AI tools, tested end to end on a practice client.
Practice case. Prairie Ledger CPA is a made-up 14-person accounting firm in Red Deer. The call transcript, client replies, records and numbers are all fictional. The skills, the method and the lessons are real.
The five skills
Each skill is a set of instructions Claude follows the same way every time. Each one hands a written document to the next.
| Skill | What it does | What it hands on |
|---|---|---|
| Discovery Lead | Runs the discovery call and finds the workflow problem behind the feature request | Discovery Brief, and a follow-up email for missing numbers |
| Value Analyst | Puts a yearly cost on each problem, checks the build is possible, sizes the fix and prices it on value | One-page proposal with pass rates |
| Prototype Builder | Builds the smallest working fix and tests it against pass rates written before testing | Build Record with the full test log |
| Trainer | Sets up the workspace, writes usage rules, plans short sessions on the team's own work | Training Plan, usage rules, cheat sheets |
| Delivery Lead | Handover, monthly value reports, scope changes and the case study | Handover summary and value reports |
How the case ran
-
Discovery
The owner asked: "I want AI to write our client emails." The call found that most of those emails chase the same clients for documents every month, and that late documents cause a late close, rushed reviews and coding errors found at year-end. Three problems were scored and carried forward. Missing numbers went into a short follow-up email instead of guesses.
-
Value and proposal
Problem Cost per year First-year value of fix Chasing documents $19,200 estimated $9,600 Repeat client questions $62,920 estimated $15,730 Coding errors $7,730 estimated $3,865 Total $89,850 $29,195 Price: $6,500 fixed, about 22% of first-year value, paying back in 2.7 months. The readiness check found that the firm's practice software already had automatic reminders, turned off. Problem 1 needed no AI at all.
-
Prototype and testing
Client Answers (Claude Project). Drafts replies to client questions using only answers the firm's accountants approved. Round 1 failed: 11 of 14 drafts needed only light edits, against a target of 12. Three common follow-up questions had no approved answer. The test design was also wrong: the original pass rate counted 6 emails the tool is meant to refuse. Round 2 passed after adding three library topics and correcting the test.
Monthly Coding Check (Claude skill). Flags transactions that may be miscoded, for the bookkeeper to decide. Round 1 caught all 10 planted errors but flagged normal food orders as large purchases. Narrowing that check opened a new gap: a $3,000 generator from a hardware store would slip through. A new check, "unusually large for this payee," closed it. The full rerun caught 11 of 11 errors with no false alarms.
Client Answers Project, test plan, round 1, round 2, coding check skill, coding check test log, reminders checklist, Build Record
-
Training
Six short sessions in December and January, before tax season. Everyone does a real task from their own week. Firm accounts and signed usage rules come before any client data is used. Each staff concern gets a plain answer, including "will this replace me?" and "it's a fad." Use is measured at 30 days.
-
Delivery
After launch, the Delivery Lead produces the handover summary, a monthly one-page value report against the proposal's numbers, and handles requests for extra work as separate scope changes. In this practice case, there are no 30-day results yet.
What testing taught me
Running the whole case end to end found gaps that no single skill would have found. Each one went back into the skills.
- Check existing software before pricing.The proposal nearly priced an AI fix for something the client's software already did. The Discovery Lead now asks which features are in use, and the Value Analyst checks before choosing a fix.
- Get every rate on the first call.Two of three problems could not be priced without the accountant rate. The Discovery Lead now collects cost and billing rates for every role involved.
- A pass rate must be possible.A target that counts cases the tool is meant to refuse can never be met. The Value Analyst and Prototype Builder now both check this before testing.
- Value a freed hour once.A freed hour is worth its cost if saved, or its billing rate if refilled with paid work. Never both. This rule is now written into the Value Analyst.
- Rerun everything after a fix.The full rerun after the coding check fix found a second problem that a partial check would have missed.
All files from the case
- Discovery Brief
- Follow-up email and client reply
- Proposal
- Problem 1: Karbon reminders setup checklist
- Problem 2: Client Answers Project
- Problem 2: Test plan
- Problem 2: Test log, round 1
- Problem 2: Test log, round 2
- Problem 3: Monthly coding check skill
- Problem 3: Test data
- Problem 3: Test plan and answer key
- Problem 3: Test log, rounds 1 to 3
- Build Record
- Training Plan and usage rules
- Cheat sheets
The five skill files will also be published on GitHub.