benchr
Not a prompt dump. Every entry names what goes in, what the move is, and what comes out different because of it.
Nothing matches that. Clear the filters to see everything again.
Instructions placed after a long attachment are followed more consistently than the same instructions placed before it. Recency is doing work that emphasis is not.
One paragraph, in its own words, before it starts. You find out whether the constraints landed while it is still cheap to correct.
A thirty-second check that tells you whether the model actually has the whole document before you trust a summary of it.
Write the output schema before the instruction. The prompt then only has to describe the judgement, because the shape is already enforced.
Split the job before the output limit splits it for you, at a boundary you chose rather than mid-sentence.
Two moves, in order of reliability: quote a short string exactly, or generate the composition with deliberate empty space and set the type in an editor.
Stop describing the subject and start re-supplying it. Then change exactly one thing per generation so you can tell what caused a drift.
A shot description that fixes camera, lens, movement, light and the constants is doing the job a shot list and a cinematographer would do. Short prompts are why most attempts look like stock footage.
Put the build command, the test command, the untouchable files and the definition of a complete change in a file at the repository root. Agents read it. So do new humans.
Your own tasks, split in two. You iterate against one half; the other half is only ever run, never tuned against. The second number is the real one.
The single highest-value substitution in analysis work. A CSV removes an entire measured failure mode that no prompt fixes.
Find the largest block of your prompt that is identical on every call — the system prompt, the style guide, the reference document — and stop paying full price for it every time.
benchr