FG-0511 Instruction Design
Description
The craft of writing instructions that produce reliable, consistent agent output. Instruction design is the discipline the field calls prompt engineering, practiced as a repeatable specification rather than a one-off query. It is the most fundamental execution skill in human-AI collaboration and the one most frequently neglected. Effective instructions are specific about desired outcomes, explicit about constraints, clear about format expectations, and unambiguous about scope boundaries. The quality of agent output is bounded by the quality of the instructions that produced it.
Implementation
Structure instructions with explicit sections: objective (what to produce), constraints (what to avoid), format (how to structure output), scope (boundaries of the task), and success criteria (how to evaluate completion). Test instructions by verifying that a fresh agent with no prior context can execute them correctly. Iterate based on output quality. When output is wrong, examine the instruction before blaming the agent.
Success Indicators
- Fresh agents produce correct output from instructions alone without clarification
- Instructions are reusable across sessions and model versions
- Output quality is consistent across repeated executions of the same instruction
- Instruction failures are diagnosed and improved systematically
Failure Modes
- Implicit context: instructions assume knowledge the agent does not have, producing output that looks correct but is wrong
- Over-specification: instructions so detailed they prevent the agent from applying judgment where judgment would improve output
- Under-specification: instructions so vague that output quality depends on interpretation rather than instruction clarity
- Static instructions: never revised despite consistent output quality issues, treating bad results as agent failures rather than instruction failures