Ibrahim Sow

I work adversarially. Every template I publish arrived with a documented failure, and I keep that failure next to it so you can decide whether the template addresses your case or merely resembles mine. Structure does more work than most writing suggests. Delimiters around untrusted content, an explicit statement of the task, and a description of the output format are the three elements that recur across templates that hold up. What does not recur is length. Padding a prompt out with restatement tends to dilute the instruction rather than reinforce it. Examples work when they show the decision you want made. Three demonstrations of the same judgement will outperform thirty showing the format without the judgement, because the model is matching the mapping from input to choice rather than the shape of the answer. Output format deserves to be pinned down rather than hoped for. If the output must parse, ask for the structure and validate it rather than accepting whatever comes back. If the output must be prose with a particular shape, describe the shape in the prompt rather than hoping the model infers it. I cover the failure modes honestly, because most of them are invisible until something breaks. Instructions buried in the middle of a long prompt are followed less reliably than the same instruction at the top. A prompt that works on a strong model and fails on a small one usually depends on capability the small model lacks. I also treat a prompt as an attack surface. Anything copied into a prompt from a user, a document or a web page is input, not instruction, and a template that does not mark where its own instructions end is a template waiting to be redirected.