01 / Understand
Make the assumptions visible
Provide the rule document and field definitions. Ask for conflicts and hand-checkable examples before implementation.
A 30-minute working session
From policy
to prototype.
Turn domain knowledge into a tool you can run, inspect, and improve with Codex.
For data analysts & developers working with complex rules.
01
Bring the business contextDefinitions · approved rules · known examples
02
Ask for working artifactsCode · checks · explanations
03
Verify the resultReview changes · test boundaries · own the decision
00–03Why AI?Start with the work
03–07Pick a taskChoose an artifact
07–11Give contextDefine “done”
11–21Build & iterateA working demo
21–25VerifyEvidence matters
25–28PrinciplesKeep accountability
28–30Your turnTry one change
The working artifact / 10-minute demo
Fictional tariff rules lab
Fictional training data — not for customs filing. Region means origin. All categories, authorities and rates are invented.
Change a shipment. See what applies, what is skipped, and what changes when a new rule takes effect.
Read the rules & resolution contract
Validate → match scope and dates → resolve each rate group → apply targeted exemptions → sum → round once.
| ID / source | Scope | Action / priority | Rate |
|---|---|---|---|
| B-S · Training Schedule §1 | Sensor / A or B | Base / 10 | 5% |
| B-T · Training Schedule §2 | Textile / A or B | Base / 10 | 8% |
| P-B · Training Preference §1 | Sensor / B | Replace base / 20 | 2% |
| S-A · Training Surcharge §1 | Sensor / A | Add surcharge / 10 | v1: 10%; v2: 12% from July 1, otherwise 10% |
| X-A · Training Exemption §1 | Sensor / A / verified certificate | Exempt S-A only / 100 | Base duty remains |
| C-S · Training Conflict §1 | Sensor / A or B / toggle enabled | Competing base / 10 | 7% |
All rules are limited to January 1–December 31, 2026, inclusive. Only the highest-priority matching rule in each rate group wins. A tie at the winning priority means REVIEW. Priority does not compare unrelated groups or establish a legal hierarchy.
USD, ad valorem only; one shared value base. Integer cents and basis points, final total rounded half up to cents. No fees, quotas, quantity duties, or real-world classification. Unknown inputs and out-of-range dates require review.
The reusable method
Give Codex a clear handoff.
02 / Build
Ask for something runnable
Specify the artifact, scope and acceptance criteria. Let Codex inspect files, implement the change and run relevant checks.
03 / Verify
Inspect evidence
Review the file changes and expected values. Try an effective-date boundary and a case that must refuse to calculate.
A prompt to start with
Read demo-brief.md. Build a self-contained index.html with Chinese as the default language and an English switch using only the supplied fictional rules. Show v1/v2 results, applied and skipped rules with reasons, and REVIEW for unresolved inputs or conflicts. Use no external dependencies or API calls. Verify the acceptance examples. Report the checks you actually ran and any remaining limitations.
Keep the judgment. Accelerate the implementation.
Four working principles.
01 · Approved data & environment
Use company-approved tools and accounts. Start with synthetic data. Local file access does not mean inference stays on your machine.
02 · Verified rule sources
Have authorized people confirm classification, applicability and effective dates. Preserve sources, versions and review records.
03 · Limited permissions
Work in an isolated project. Keep production writes, deployments and external sharing behind explicit authorization.
04 · Reproducible evidence
Keep inputs, rule versions, traces and known-answer checks. Escalate ambiguity. A confident explanation is not proof.
Your turn: What recurring task could you pilot with synthetic data this week?
Measure time to a reviewed result, corrections needed, and known-case accuracy.