Rapid Open-Goal Design Lessons after IU Core 240x

Date: 2026-05-22 Status: lesson note for future projects

Context

IU Core / Miếng thông tin moved much faster after switching from small checklist prompts to large open-goal macros. The system reached production-grade substrate, runtime, structure ops, delivery seam, three-axis metadata, piece-native composer, one-command operations, and text-as-code kickoff through progressively larger prompts: 30x → 60x → 120x → 240x.

Main lesson

For capable coding agents, the expensive part is usually reading context, understanding architecture, and setting up the execution frame. Once that cost is paid, the actual implementation/proof/reporting may be much faster than expected.

Therefore, do not split related product work into tiny prompts. Design a large coherent outcome and let the Agent own the path.

What a good design should contain

A fast implementation design should be short enough to maintain but complete enough to execute:

1. Desired end-state, not micro-steps.
2. SSOT and live survey requirement.
3. PASS / PARTIAL / BLOCKED criteria.
4. Forbidden list.
5. Gate rules for production mutation.
6. DOT / no-hardcode / five-layer requirements.
7. Discover-first and automation-first rules.
8. Evidence and report verification.
9. Next macro package.

What to avoid

• Over-detailed step-by-step prompts.
• Readiness-only mini tasks when implementation is safe.
• Repeating context in many small prompts.
• Creating design documents that lag behind implementation.
• Treating git commit as KB/AgentData upload.
• Allowing manual-memory procedures instead of DOT/script/command packages.

Recommended pattern

Use a high-scale open-goal prompt when work shares the same product axis.

Current experimental scale after IU Core 240x: 480x may be tested next if 240x is verified and no safety debt appears.

Scale must not weaken safety. Larger prompts require stronger evidence, not weaker gates.

Non-negotiables

• live survey before action;
• no hardcode;
• DOT 100%;
• five-layer sync;
• discover-first/reuse-first;
• automation-first / one-command operations;
• reversible by default;
• production mutation gated;
• AgentData upload/list/read/search before PASS;
• no fake PASS.

Design update rule

After a major implementation burst, update the roadmap/design only enough to keep the spine accurate. Do not spend days perfecting a design that the Agent can validate by implementation in one macro. The design should guide the next large mission, not become a bottleneck.