hackathon-project-verification-patterns

name	hackathon-project-verification-patterns
description	Patterns observed in Build Day projects for planning, verification with independent sub-agents, and cost/latency optimization under tight time constraints.

Instructions

Use these patterns when you need to build a working demo quickly, while still keeping outputs verifiable and costs under control.

1. Plan before building

• Write a lightweight PRD with success criteria.
• Break work into parallel workflows (e.g., data prep, UI, backend, evaluation).

2. Add an evidence chain (when claims must be traceable)

• For each generated component, record where it came from (source document, image, diagram) and why it was placed/derived.

3. Verify with independent sub-agents

• Run verification in isolated context windows so reviewers are not biased by prior steps.
• Use a self-correction loop that rechecks failed components until all tests pass.

4. Optimize expensive simulations

• If you start with per-entity calls (e.g., one call per persona/agent), look for batching or clustering approaches that preserve accuracy.
• Evaluate against real-world baselines (historical outcomes, known distributions, external benchmarks).

5. Use Claude for tool selection, not just code writing

• Ask Claude to research and choose the right model/tool for a specific subtask.
• Hand it unfamiliar technologies and ask it to integrate them into your pipeline.

Examples

Example: verification loop with independent reviewers

See examples/verification-loop.md.

Example: clustering to reduce inference cost

See examples/cost-cutting-clustering.md.

Source

• https://claude.com/blog/meet-the-winners-of-our-claude-opus-4-8-build-day-hackathon