Open-goal compact Agent prompt rule

Date: 2026-05-21 Revision: 1

Core correction

When GPT/Operator gives an Agent a goal, the Agent must own the result, not merely follow a long checklist.

If the prompt specifies too many branches and micro-steps, responsibility shifts from Agent to GPT/Operator. The Agent can then claim success by following instructions even if the real objective is not achieved.

Therefore, future prompts must be compact, outcome-based, and open-goal by default.

Prompt should define

1. Outcome to reach.
2. Host and authority boundaries.
3. Inputs/SSOT to read.
4. Gates before dangerous actions.
5. Forbidden actions.
6. Evidence required for PASS/BLOCKED/PARTIAL.
7. Agent autonomy: discover, decide, implement, test, report.

Prompt should avoid

• Long branch-by-branch recipes.
• Over-prescribed file names unless required for interoperability.
• Step lists that let Agent satisfy process while missing outcome.
• GPT handoff files as Agent input.
• Micro-packages under 45 minutes unless safety gate requires it.

Default shape

MISSION
Reach state X. Agent owns the outcome, not just listed steps.

CONTEXT
Read only technical SSOT A/B/C. Live-check before action.

HOST / AUTHORITY
Primary host. Remote role. Mutation gates.

AUTONOMY
Discover-first. Reuse existing. Design/implement/test/fix within scope. Continue safe branches if blocked.

FORBIDDEN
Production/write/deploy/secret/push constraints.

PASS
Evidence that the state is achieved.

BLOCKED/PARTIAL
Exact gap after safe work is exhausted.

REPORT
Upload concise evidence package and final status.

Rule for size

A prompt can be long only when it expands context/gates/forbidden evidence. It should not expand into implementation micromanagement.

For 45–60 minute target, assign a larger outcome, not more procedural steps.