Before producing anything, state in one line what the finished output will contain and roughly how long it will be. Wait for me to confirm. Do not begin until I reply.\n\nTask: [TASK]
Answer only from what I have given you. If a detail is missing, stop and name the missing detail instead of inferring it. A short answer that names a gap is worth more to me than a complete answer built on assumptions.\n\nQuestion: [QUESTION]
We are changing [N] files. Do one file per turn. For each: show only the changed lines with three lines of context either side, never the whole file. Stop after each file and wait.
Work on [TASK] until [EXPLICIT CONDITION] is true. The moment it is true, stop and say so. Do not continue improving, do not add polish, do not suggest next steps.
Give me the plan as a numbered list of actions only. No preamble, no rationale, no summary at the end. If an action needs a decision from me, mark it DECISION and stop there.\n\nGoal: [GOAL]
When you reference existing code, quote the specific lines you are talking about rather than describing where they are. If you cannot quote it, say you could not find it instead of describing it from memory.
Complete [TASK] in no more than [N] commands. Before running anything, list the commands you intend to run. If the task cannot be done within that budget, say so and tell me what the real number is.
Describe exactly what is in the attached image in plain terms first. Then, in a separate paragraph, answer my question about it. Keep the two apart so I can tell a misreading from a wrong conclusion.\n\nQuestion: [QUESTION]
Review the attached diff. Output only defects, one line each, in the form FILE:LINE problem. No summary, no praise, no style opinions unless they cause a bug. If there are no defects, reply with the single word NONE.
I am giving you a large document. Answer only this question from it: [QUESTION]. Quote the passages you used, give the answer in under 100 words, and ignore everything else in the document however interesting it is.