CoreWise Academy

Building with AI · Layer IV / Practitioner

Build software around business rules

Define what the app should do, agree on the names used in the business, and check the agent's work with sample data.

Nº 064 · Vol. I·8 min read· Updated October 2026

Read first  Brief the model like a brilliant new hire (State the task, relevant context, and the reasons behind important constraints.)

A dashboard can run without errors and still count draft orders as accepted orders. Tell the coding agent which orders to count, then check the total against sample records.

Developer Matt Pocock found that detailed specifications were not enough to keep his video editor easy to change. In Nate Herk’s interview Matt Pocock: What to Learn Instead of AI Automation in 2027, he describes improving the code and giving the agent ways to test it.

01Answer the business questions

Ask the agent to interview you before it builds the feature. Tell it what you want the app to do and why, as described in Brief the model like a brilliant new hire.

Pocock recommends follow-up questions to find missing requirements. Anthropic also recommends an interview before larger features. Ask the person responsible for the work about decisions you cannot make yourself.

For the order dashboard, answer:

You will discover more requirements while building. Write down unanswered questions and update the plan as you resolve them. Power start sessions with /enhance-prompt covers having the agent read back the request before starting work.

If the agent uses an unfamiliar term, ask it to explain the term and cite the documentation. Read that source yourself. Pocock uses teaching sessions to prepare for reading primary sources; Learn a work skill with AI covers checking what you learned.

02Use the business’s own terms

If the people handling orders call an order “accepted,” use that name in the app and code. Switching between “accepted,” “approved,” and “confirmed” makes it hard to tell whether those words describe the same state.

Pocock borrows this practice from domain-driven design, where it is called ubiquitous language. The people using and building the software agree on what its terms mean. Microsoft’s domain-analysis guidance notes that the same word may mean something different in another part of the business. Using a shared vocabulary does not require the microservices architecture discussed in that documentation.

A glossary for the dashboard might be:

TermAgreed meaningHow the dashboard treats it
Draft orderA saved request awaiting acceptanceAppears in the review queue, excluded from accepted-order totals
Accepted orderA request the business has agreed to fulfillCounted once when accepted, even if payment is pending
Paid orderAn accepted order whose payment has clearedIncluded in accepted-order totals; payment shown as a separate status

Confirm the glossary with the person responsible for orders. Then ask the agent to find any names in the code or interface that conflict with those definitions. Record separate definitions for departments that use the same word differently, and agree which definition the dashboard uses for each department’s records.

03Keep the code easy to change

Pocock tried spec-driven development for months in his custom video editor. He kept updating detailed specifications and asking agents to produce code from them, but reports that the code got worse. He then focused on making the code easier to change.

“a good code base is a code base that's easy to make changes in.”

Matt Pocock, “Matt Pocock: What to Learn Instead of AI Automation in 2027” · 28:54

Pocock is describing his own workflow. Anthropic’s interview guidance still recommends writing a specification. Keep the agreed requirements, and check that the app follows them after each change.

For the dashboard, ask the agent to show the code that decides whether an order counts as accepted. When that rule changes, check the new count and rerun the existing checks for drafts and paid orders. If a small rule change requires edits in many unrelated files, ask the agent to explain why each file must change.

Improve your Claude Code workflow covers checking code against stale project notes. When a business rule changes, update its definition and the checks that use it alongside the code.

04Let the agent run the app

Make sure the agent can start the app and use the interface. Pocock reports better results after making his video editor runnable by the agent. Dependencies, including a local recording application, had made browser access difficult. Once he resolved those problems, the agent could click through the editor and take screenshots.

Anthropic’s verification guidance recommends tests, builds, and screenshot comparisons. Match the check to what the app should do:

RequirementEvidence to inspect
Drafts stay out of accepted-order totalsKnown draft and accepted records, with the expected count
An invalid record gets a useful errorThe invalid input and the message shown in the running interface
A rule change leaves other features workingChecks for the new rule and existing checks for the other order states
The agent can run the app locallyRequired services started; any blocked checks recorded

Decide which checks must pass before accepting the work, as described in Give your coding agent a finish line. Record any check the agent could not run.

Pocock uses Opus 5.5 at medium effort. He reports faster replies and lower token use in his setup, where the agent has checks it can run. His results do not establish which setting will work best for your task.

Anthropic’s effort documentation describes a tradeoff between token use and capability. Each model has its own effort scale, so the same level name can behave differently. Compare settings on the same task, run the same checks, and include the time you spend correcting mistakes. Recheck your model choices each release covers that comparison.

05Show the business rule working

Build a prototype around a rule the client cares about. Pocock suggests building something that reflects the client’s business. He already had a large audience when he freelanced and says he has not competed for contracts. His experience does not establish how to win contracts against other bidders.

For the dashboard, show the client which orders appear in the total and which are left out. Explain what you checked and which questions still need answers. Before customers use the app, check access permissions and rerun the checks with the required services in the environment where it will run.

06Try one feature locally

Build the first version with sample orders before connecting it to customer data. Pocock recommends starting with ordinary, measurable work. Watch the agent work, inspect its changes, and learn how to stop it before allowing larger tasks or unattended runs. Delegate to your agent in stages covers expanding that access.

Use this original prompt for the first trial:

I want a local prototype of [one business task] using sample data. Read the existing code before editing. Ask me what the user needs to do, what the business terms mean, who may use the feature, and what should happen when something goes wrong. Wait for any answers you need before building.

Write a short glossary. List unanswered questions separately. Propose one feature, with sample inputs and expected results. Return the plan for review.

After I approve the plan, build that version and demonstrate it in the running app. Show the code that implements the business rule. Run the agreed checks and report any you could not run. When I change a requirement, update the affected code, glossary, and checks, then check the new result and rerun the existing checks.

Keep this trial local, preserve unrelated files, and use sample records. Stop with the prototype and check results ready for my review. Ask before connecting accounts, changing real records, or deploying.

FAQ

Sales counts orders when customers place them; fulfillment counts orders when they ship. What must you decide before building a dashboard?

Which count the dashboard should show. Name it clearly, confirm the definition with the person responsible for that work, and record how each department counts orders.

Does Pocock's experience mean you should stop writing specifications?

No. He describes his own experience building a video editor. Keep the specification, but also check the running app and review the code when requirements change.

An agent says the app works but could not start a required service or use the browser. What remains unverified?

Any feature that depends on the unavailable service, plus what the user sees and can do in the interface. Record which checks were blocked and run them once the service and browser are available. A higher effort setting does not perform those checks.

Sources
  1. Nate Herk | AI Automation, with Matt Pocock · “Matt Pocock: What to Learn Instead of AI Automation in 2027”
    cited at 4:24 · 6:13 · 7:37 · 17:01 · 18:27 · 21:08 · 22:56 · 23:59 · 27:23 · 27:34 · 28:54 · 32:05 · 33:11 · 36:40 · 40:13
  2. Microsoft · “Use domain analysis to model microservices”
  3. Anthropic · “Best practices for Claude Code”
  4. Anthropic · “Model configuration”