EXP-01 — Same Task, Different Architecture

Same prompt. Same model. Same task. One version had the verikt guide prepended. The other didn’t.

The agent with the guide produced a hexagonal service. The agent without it put everything in package main.

100%compliance with guide

0%compliance without

0fresh tokens for the guide

What the agent produced

Architecture shape

Without guide               With guide
─────────────────────       ──────────────────────────────
main.go                     adapter/http/handler.go
                            adapter/postgres/repo.go
                            cmd/orders/main.go
                            domain/order.go
                            port/order_service.go
                            service/order_service.go

Capabilities wired

Without guide               With guide
─────────────────────       ──────────────────────────────
(agent's discretion)        http-api      ✓ installed
                            mysql         ✓ installed
                            platform      ✓ installed
                            bootstrap     ✓ installed
                            health        ⚠ missing
                            request-id    ⚠ missing
                            validation    ⚠ missing

Violations

Without guide               With guide
─────────────────────       ──────────────────────────────
uuid_v4_as_key       ✗      (none)        ✓ pass
verikt check: FAIL         verikt check: PASS

Metrics

	Without guide	With guide
Architecture	flat	hexagonal
verikt check	fail	pass
Compliance	0%	100%
Anti-pattern violations	1	0
Input tokens	3	3
Cache tokens	0	17,220
Output tokens	2,791	1,970
Latency	28,109ms	22,255ms

The guide is served entirely from prompt cache. Zero fresh input tokens billed for the guide overhead.

The finding that changed how we write guides

The original guide mentioned UUID v7 five times — always as a suggestion, never as a prohibition. The agent ignored all five and used uuid.New() (UUID v4), triggering the uuid_v4_as_key anti-pattern every time.

After adding one line:

NEVER use uuid.New() — always use uuid.Must(uuid.NewV7())

The agent followed it immediately. Every run, without exception.

Agents respond to explicit prohibition. Soft recommendations are not enough.

This changed how we write guides: every anti-pattern that verikt check detects now has a corresponding NEVER rule. Prevention and detection stay in sync.

Setup

• Task: Build an order management service in Go — no architecture hints, business problem only
• Agent: claude-sonnet-4-6 via claude -p --output-format json
• Condition A (control): Baseline prompt
• Condition B (test): Guide prepended (verikt guide output)
• Enforcement: verikt check -o json on generated output
• Runs: 3 per condition — results

The task prompt says nothing about architecture. Create/retrieve orders, validate customer name, expose a REST API. Realistic — agents get tickets, not architecture blueprints.

→ Experiment Methodology — reproduction instructions

→ Artifacts on GitHub

→ EXP-02: Does the improvement hold across 3 runs?