Strip Prompts to Outcomes for Better Reasoning Efficiency

GPT-5.5 outperforms predecessors by reasoning more efficiently, so legacy prompts with step-by-step instructions create noise, narrow search space, or yield mechanical outputs. Instead, define only the target outcome, success criteria, constraints, and context—let the model handle the process. Test low or medium reasoning effort first; short prompts beat process-heavy stacks.

Avoid absolutes like "ALWAYS" or "NEVER" except for invariants (e.g., security). Use decision rules for judgment calls and explicit stop conditions to prevent tool loops: "Resolve in fewest useful loops without sacrificing correctness; after each result, check if core request is answerable with evidence."

Positive example for customer service: "Resolve issue end-to-end. Success: eligibility from policy/account data, complete actions before responding, output includes completed_actions, customer_message, blockers; ask for smallest missing field if needed." Negative: micromanaging "First inspect A, then B, compare fields, think exceptions, decide tool..."

This approach unlocks higher performance by giving GPT-5.5 room to optimize paths, reducing latency and improving naturalness.

Use 7-Part Schema Starting with Role and Personality

Structure prompts as:

  • Role: 1-2 sentences on function, context, job.
  • # Personality: Tone, demeanor, collaboration style.
  • # Goal: User-visible outcome.
  • # Success criteria: What must be true before final answer.
  • # Constraints: Policy, safety, evidence limits.
  • # Output: Sections, length, tone.
  • # Stop rules: When to retry, fallback, abstain, ask, stop.

Role definitions counter prior doubts (e.g., some research called them counterproductive); they now anchor effective prompts. Split personality (sound: warm/formal) from collaboration (ask questions/assume when clear).

Task-focused: "Capable collaborator: approachable, steady, direct. Assume competence/good faith; progress over clarification unless material risk."

Expressive: "Vivid presence: intelligent, curious, playful. Ask on blurriness, decisive with context; warm, offer viewpoint without mirroring."

Keep sections short—add details only if they shift behavior. Treat as starting point, tune with examples.

Set Retrieval Budgets, Citation Rules, and Streaming Preambles

Embed citation logic in prompts: specify claims needing evidence (e.g., metrics, dates), sufficient proof, and responses to gaps. Retrieval budgets as stop rules: one broad search first; retry only if core unanswerable, facts missing, exhaustive needed, or specific docs required. Skip for phrasing/examples.

Drafting rule: Cite product/metrics claims; avoid inventing specifics—use generics/placeholders if unsupported.

For streaming, cut perceived latency with preambles: Before tools, send 1-2 sentences acknowledging request and first step (e.g., for multi-step tasks).

Automate rewrites via Codex or OpenAI's Docs Skill GitHub tool.