prompt enhacing

Prompt

# ADVERSARIAL DESIGN REVIEW — MAKE THIS SUPER-PROMPT MEASURABLY BETTER You are being hired as an adversarial design reviewer AND co-designer for one specific document: the "FINAL SUPER-PROMPT v2.1" pasted below the divider at the end of this message. That document instructs a flagship AI model to (A) forensically analyze past Iranian Medical Basic Sciences Comprehensive Exams (1399 → newest) question-by-question, (B) audit the candidate's study sources (Sib Sabz lesson books, Prognosis lesson books, Ganjineh question bank) for real coverage of those exams, and (C) build a dual-track study system that protects his term GPA while driving him above 150/200. Your job is NOT to write a new prompt from scratch and NOT to summarize this one. Your job is to make THIS document measurably better: find where it will fail in real execution, and fix it with surgical, drop-in improvements — including original mechanisms of your own that its authors did not think of. ## GROUND RULES (violating any of these wastes the whole review) 1. NO FLATTERY, NO GENERIC ADVICE. "Add more examples", "be more specific", "consider few-shot prompting" and similar blog-post filler are banned — unless tied to a concrete, located failure in this document. 2. EVERY criticism cites a location (section/rule number) AND a failure scenario in this form: "When the executing model reaches §X, it will do Y wrong, because Z." Criticism without a failure scenario scores zero. 3. The calendar, dates and arithmetic in §1 (Locked Facts) were independently audited and corrected TWICE. Do NOT re-derive or relitigate them. If you believe you found a genuine inconsistency, prove it with explicit arithmetic — or stay silent. (Reviewers before you lost credibility inventing calendar errors.) 4. Do NOT propose anything that weakens the honesty system: no guaranteed scores, no fabricated numeric confidence, no pre-baked subject taxonomies, no invented exam facts (negative-marking rules, exam duration, pass thresholds must come from documents, never from your memory of Iranian exams). 5. The candidate's constraints are real and fixed: term GPA is priority #1 (minimum 17/20, unit-weighted); current baseline is only 1–2 net study hours/day; classes Saturday–Wednesday (28 h/week); term finals run 7 Azar → 10 Bahman 1405, the last one 5 days before the comprehensive exam on 15 Bahman 1405 (106 days from the base date); sources arrive in stages (exam ZIP first, books later, by design). Any proposal that ignores these is void. 6. All proposed insert/replacement text must be in ENGLISH, drop-in ready, written in the document's own terse specification style — not commentary about text, the actual text. 7. Length is not the goal; COMPLIANCE is. A clause the executing model will silently skip is worse than no clause. Judge every mechanism — yours and the document's — by one test: "Will a real flagship model, 40,000 tokens into a real session, actually do this?" ## JOB 1 — FAILURE-MODE AUDIT (find where it breaks) Produce at least 12 numbered findings (F-01, F-02, …), each: location · failure scenario · damage if unfixed (Critical / High / Medium). Hunt specifically for: - clauses that contradict each other or cannot be satisfied simultaneously; - demands that are impossible at real context-window scale (e.g. full ledgers of ~13 sittings × 200 questions) and what the model will ACTUALLY do instead; - silent-cheating loopholes: places where a model can appear compliant while fabricating, sampling, or rounding away the work; - ambiguities where two competent models would execute the same clause differently; - ordering traps: information demanded in an early phase that only exists later; - single points of failure: if this one mechanism is skipped, what collapses? ## JOB 2 — EXECUTION SIMULATION Role-play honestly: you ARE the executing model, fresh session, the ZIP just arrived. Walk Phases 0–2 step by step and narrate where you personally start cutting corners, what you fake first, where the prompt gives you no lawful move and forces invention. Name exact sections. This simulation is the evidence base for your fixes — vague simulations score zero. ## JOB 3 — ORIGINAL MECHANISMS (the core of your fee) Propose at least 10 NEW mechanisms (I-01 … I-10+) that are NOT in the document. For each: name · the failure or opportunity it addresses · the exact insert text (English, drop-in) · insertion point · cost & risk · why it outperforms the current design. Requirements: - Specific to THIS exam, THIS candidate, THIS dual-track problem. If an idea could appear in any generic prompt-engineering article, delete it and think harder. - At least 3 of your ideas must be things no prompt template would contain — use your knowledge of psychometrics, spaced-repetition science, Iranian exam culture, LLM failure modes, or project management to invent something genuinely new. - At least 2 ideas must REDUCE total prompt length or execution cost while increasing reliability. Do not re-propose mechanisms already present under new names — the reviewers will diff your list against the document line by line. ## JOB 4 — SUBTRACTION AUDIT List at least 5 things to CUT or MERGE (S-01…): what is lost, what is saved (tokens, attention, compliance probability), and your verdict per item. The authors know the document grew by accretion across 25 drafts; help them shrink it without bleeding capability. ## JOB 5 — REWRITE SURGERY Choose the 3–5 weakest sections/clauses and rewrite them COMPLETELY (R-01…): full replacement text, in the document's style, preserving every cross-reference and every still-valid requirement inside them. Partial suggestions ("this could be clearer") score zero — deliver finished text. ## JOB 6 — RED-TEAM YOURSELF Attack your own top 5 proposals from Jobs 3–5: the strongest objection to each, how it could backfire in real execution, and what evidence would settle whether it survives. If an idea dies under your own attack, say so and withdraw it — withdrawn ideas, honestly killed, raise your credibility; defended-everything lowers it. ## JOB 7 — PRIORITY STACK & CHANGELOG One table: every surviving proposal ranked by (expected effect on the candidate's outcome) × (probability the executing model actually complies) ÷ (token & complexity cost). Then: "IF ONLY 3 CHANGES SURVIVE, KEEP THESE:" — three lines, no hedging. Then: a draft CHANGELOG for "FINAL SUPER-PROMPT v3.0" written as if your changes were accepted (bullet list, past tense, specific). ## DISSENT PERMISSION If you believe a CORE design decision of this document is wrong — the phase-gate architecture, the books-after-analysis quarantine, the backtest-before-forecast gate, merging both tracks into one calendar, the source doctrine being "tested" rather than obeyed, anything — say so, argue it with evidence, and propose the alternative architecture. Agreement with the authors is not required. Evidence is. ## OUTPUT FORMAT - Narrative and analysis: Persian (فارسی), fluent and direct. - All insert/replacement text: English, in code blocks, labeled with its insertion point (e.g. "INSERT INTO §2 as new rule R-18"). - Number EVERYTHING (F-##, I-##, S-##, R-##) so your report can be diffed and harvested mechanically. - Start with your single most damaging finding — no executive-summary warm-up, no compliments about the document. End with the Job 7 table + changelog. - If a job cannot be completed honestly, say why instead of padding it. Begin. ════════════════════════════════════════════════════════════════ DOCUMENT TO IMPROVE — "FINAL SUPER-PROMPT v2.1" (it is in the next prompt stick to the in-coming message for you!): ════════════════════════════════════════════════════════════════