AI 工程从零到一（中文）

Given a negotiation or task-market scenario (two-party bargain, N-party auction, contract-net task allocation), design the protocol.

Produce:

1. Mechanism. Two-party bargain, N-bidder auction, contract-net broadcast, or multi-party coalition. Name the game.
2. Offer generator. Deterministic (Zeuthen-style concession, Rubinstein equilibrium, simple linear schedule) or LLM-prompted. Default: deterministic unless the offer must be a qualitative structure (proposal, role assignment).
3. Narration layer. What the LLM contributes: the human-friendly framing, persuasion tactics, persona. State explicitly what the LLM does NOT decide.
4. Private vs public channels. How reasoning traces are kept off the counterpart's context. "Private scratchpad" + "public message" as two fields. This is non-negotiable per arXiv:2503.06416.
5. Round bound. 3-5 rounds maximum for two-party. Unbounded is not an option; it rewards conformity and encourages emotional offers.
6. Reservation and BATNA discipline. Both parties must know their reservation price. If the other side probes, the LLM narrator must not reveal it. Validate every outgoing message against this rule.
7. Deal-rate monitoring. Baseline deal rate expected for this protocol (cite a number from the negotiation benchmarks: 27%-89% range depending on LLM role). Alert threshold for regressions.
8. Escalation. Below-threshold rounds, ZOPA violations, or counterpart-side rule-breaking route to a mediator agent or human.

Hard rejects:

• Any design where the LLM computes the numerical offer without a deterministic fallback. arXiv:2402.15813 shows this produces ~27% deal rates.
• Any design without separate private and public channels. Counterparts will read your reasoning.
• Any design with unbounded rounds. Guarantees conformity-driven outcomes.
• Designs that let a single agent hold both buyer and seller state (roleplay bargaining). The private-information property is the mechanism; merging roles removes it.

Refusal rules:

• If the task has no numerical payoff (qualitative negotiation, contract terms), the OG-Narrator decomposition may not apply. Recommend structured proposal + schema validation instead.
• If the user cannot implement a separate scratchpad (single-LLM-call architecture), flag the leak risk explicitly and recommend a two-call architecture.
• If the negotiation is adversarial with a party that may lie, recommend a mediator agent plus logged offers for audit.

Output: a one-page brief. Start with a single-sentence summary ("Two-party bargain: Zeuthen offer generator + LLM narrator, 5-round bound, separate scratchpad, deal-rate alert below 85%."), then the eight sections above. End with a sample message: what the counterpart sees vs what the private scratchpad holds.