Failure-Derived Prompt Tips

Encode recurring failed strategies seen in development trajectories as short imperative tips in the instance prompt to steer later episodes.

Failure-derived prompt tips are manually iterated instructions distilled from inspecting failed development trajectories — then pasted into the instance layer of an agent episode prompt stack. Examples from software-engineering agents: always reproduce the bug before editing; never retry an identical failed command; jump with goto instead of endless scrolling; verify the open file versus cwd after create/open; re-read the file after edit.

Tips are surprisingly effective at killing the specific errant strategies they name, even when full demos do not teach domain nuance. They do not scale as a manual craft — treat them as a bridge toward automated identify-failure → write-corrective-instruction loops, closer to Reflexion but baked into the static prompt rather than only emitted mid-episode. Keep tips concrete and imperative; vague “be careful” advice under explicit instruction design rarely changes tool choice. For interactive ReAct prompts, prefer few-shot thoughts that advance subgoals over repeating the same unfinished goal each turn — ReAct-IM-style loops are a failure mode you can spot in development trajectories and ban explicitly.

---

Sources

• SWE-agent (Yang et al.), pp. 31–45