Design a summarizer prompt that preserves user-stated facts under aggressive compression

A summarizer prompt is not just "compress this conversation in 500 tokens." That instruction lets the model decide what to keep, and the model will keep whatever makes the summary read fluently. User-stated facts are usually not what makes prose fluent, so they tend to be the first thing dropped. The fix is **structure**. Tell the summarizer to produce a sectioned output, not free-form prose. A good shape has four sections: - **Durable user facts**: things the user stated that should survive across the whole conversation (allergies, preferences, identifiers, hard constraints). - **Open items**: questions or tasks the user is still waiting on. - **Completed items**: tasks already done, with their key outcomes. - **Narrative**: a brief paragraph tying the conversation together. Each section gets its own token cap. The durable-facts section, say, gets 150 tokens of its own and is forbidden from being trimmed to make room for narrative. Three instructions are essential. First, **quote user statements verbatim** rather than paraphrasing. Second, **omit polite chatter and acknowledgments entirely**. Third, **never invent** details that did not appear in the turns. The failure mode being designed against is fact loss disguised as polished prose: a beautiful summary that no longer remembers the user said "I am allergic to peanuts."

The prompt should produce a **sectioned** output rather than free-form prose, because free-form prose lets the model trade off user-stated facts against narrative polish, and polish wins. Four sections cover the working set. **Durable user facts** holds things the user stated that should survive across the whole conversation: allergies, preferences, identifiers, hard constraints, and decisions. **Open items** holds questions or tasks still pending. **Completed items** holds tasks already done with their key outcomes. **Narrative** is a brief paragraph tying the conversation together. Each section gets its own protected token budget. The durable-facts section might own 150 tokens of the 500 total and is forbidden from being trimmed when narrative wants more space. This is the key design move: without per-section caps, the summarizer trades facts for fluency at every refresh. Four instructions are essential to the prompt body. First, **quote user-stated facts verbatim** rather than paraphrase. The summarizer should copy the user's words inside the facts section, especially for numerics, names, and constraints. Second, **omit polite chatter and acknowledgments entirely**, since they contribute nothing across refreshes. Third, **flag any contradictions** the summarizer spots between turns, so the downstream model can resolve them rather than silently averaging. Fourth, **never invent details** that did not appear in the source turns. The failure mode being designed against is fact loss disguised as polished prose. A free-form summary at 500 tokens reads beautifully, scores well on semantic-similarity evals, and quietly no longer mentions that the user said "I am allergic to peanuts." Sectioning, per-section caps, and the verbatim-quote rule are the structural defenses against that failure.