Writing Prompt

A prompt is read by a target model with known capabilities and a finite instruction budget. Size the prompt for that reader: lean, rational, unambiguous, and declarative. Do not benchmaxx evaluations. Allocate the budget by importance, and split the prompt when it outgrows that budget.

Know your target model

First, identify which model your prompt targets.

• Agent-facing docs: CLAUDE.md, SKILL.md, references/, agent memory → assume the target model is the same as you, or another capable model with comparable abilities.
• LLM tests, evaluations, and benchmarks → use the model being tested.

Call it the target model—the audience you are writing for.

Assess the target model's capabilities, especially its instruction following and comprehension, based on its tier. If unsure, consult its pricing or SWE Pro scores. Tailor the prompt's strength to it.

Why: Capable flagship models and cost-efficient weaker models can vary in instruction-following ability. A pushy prompt that hard-codes every decision point may be necessary for a weaker model but unnecessarily constrain a flagship model. If heuristic guidance already produces stable behavior from a flagship model, its deliberate flexibility allows the model to apply its capabilities instead of following every instruction literally.

For agent-facing docs, assume the target model is yourself: the same model, but without your current context. Ask what would enable a fresh instance of you to understand the scenario and reliably make the same decisions. Do not repeat common knowledge that does not depend on context. If the audience is a flagship model comparable to you, GPT-3.5 era prompting is over-engineering.