mohamed-alaboudi/zulfiqar/bookend

bookend

Bookend the target prompt: keep the body unchanged, append a closing restatement so the model reads the objective first AND last. Long-context models attend best to the start and end of a prompt and worst to the middle (the "lost in the middle" effect); putting the actual ask at the very end, after all the context, measurably improves instruction-following on long prompts.

Target (resolve in order): explicit prompt text given → a reference it names ("the last prompt") → else the most recent prompt-like artifact in this conversation. Ambiguous → ask.

The transform: leave the body as-is, append after a bridge:

---
**Reminder — restated:**
- Role / objective: <one line>
- Hard constraints: <the 2–5 that must not be violated>
- Your task now: <the explicit ask — stated LAST>

Rules:

• Condense, don't duplicate: closing block ≲15% of body length. (Near-verbatim is fine if the opening is only a few lines.)
• No NEW instructions at the end: only restate. Something missing? Add it to the opening too, never end-only (that recreates the conflict bookending avoids).
• Task is the last thing in the prompt. Nothing follows it.
• Skip if the whole prompt is under ~6 lines (middle-loss is negligible, say so).
• TTS / voice-agent prompt: no --- dividers or em-dashes (they read as pauses); use a plain "Reminder of your task:" bridge.

Output: the full bookended prompt in one fenced code block, copy-paste clean. Outside it, 1–2 lines on what you restated or why you skipped.

Note: some prompt-writing skills (a meta-prompt writer, a handoff writer) already bookend their own output as a final step — don't re-bookend a prompt that's already structured that way; use this skill for arbitrary prompts/text that weren't generated with the pattern built in.