Prompt Length Optimisation

Signal vs Noise

Every word in a prompt either helps the model understand what you want or dilutes the signal with noise. Longer prompts are not inherently better — they're only better when the additional length carries genuine informational value. Pleasantries ("please kindly"), hedges ("if possible, maybe"), and redundant restatements of the same constraint add tokens without adding meaning. The model processes them, but they don't improve output quality. Precision beats verbosity.

When Context Genuinely Helps

There are tasks where longer prompts consistently produce better results. Providing background context helps when the model needs domain knowledge it might not assume: "This is a B2B SaaS product targeting enterprise procurement teams" gives the model a lens it couldn't otherwise construct. Providing the document, data, or source material you want the model to reason about is obviously necessary. Specifying edge cases and constraints that aren't obvious also earns its length. The test: would removing this sentence change what the model produces? If no, cut it.

Removing Filler Phrases

Common filler patterns to eliminate from prompts: - "Please" and "kindly" — the model isn't going to refuse because you're rude. - "I was wondering if you could…" — just ask. - "As an AI language model…" — the model knows what it is. - "Feel free to…" — this is implicit permission that the model already has. - Restating the same constraint twice in different words.

Structured Prompts for Complex Tasks

When a prompt genuinely needs to be long, structure saves it from becoming unreadable. Use markdown headers to separate sections: ## Task, ## Context, ## Constraints, ## Output Format. This lets the model parse the different components cleanly rather than trying to extract meaning from a wall of text. Structure also makes it easier to iterate — you can edit one section without touching the others.