# Prompt guidance for OpenAI GPT-5.5

> OpenAI's official GPT-5.5 prompt guide: legacy ALWAYS/NEVER prompts now hurt performance. Use outcome-first prompts with stop rules and verification loops.

Published: 2026-05-03
URL: https://daniliants.com/insights/prompt-guidance-openai-api/
Tags: openai, gpt-5-5, prompt-engineering, agents, llm, tool-use, codex

---

## Summary

OpenAI's official prompt guide for GPT-5.5 (and 5.4 / 5.3 Codex) argues that legacy "step-by-step ALWAYS/NEVER" prompts now hurt performance. Newer models prefer outcome-first prompts with explicit success criteria, stop rules, retrieval budgets, and verification loops, plus separated personality vs. collaboration-style blocks. It is the canonical migration reference for any production prompt stack moving to 5.4/5.5.

## Key Insight

- Strip legacy over-specification. Older prompt stacks pile ALWAYS/NEVER/must rules to compensate for weaker models; 5.5 reads them as noise that narrows search space and produces mechanical answers. Reserve absolutes for true invariants (safety, required output fields).
- Split "personality" from "collaboration style". Personality controls how it sounds (tone, warmth, formality). Collaboration style controls how it works (when to ask, when to assume, how proactive). Mixing them makes both blocks weaker.
- Outcome-first prompt template: Role, Personality, Goal, Success criteria, Constraints, Output, Stop rules. Each block short. The "Stop rules" block is the most underused, without it, agents loop or over-search.
- Retrieval budgets are stop rules for search, not soft suggestions. Concrete pattern: one broad search first; only re-search if (a) top results don't answer the core question, (b) a required fact/owner/date/ID is missing, (c) user asked for exhaustive coverage, or (d) a specific document must be read. Don't re-search to "improve phrasing" or cite nonessential details.
- Add a preamble pattern for streaming UX: "Before any tool calls for a multi-step task, send a short user-visible update that acknowledges the request and states the first step. Keep it to one or two sentences." Improves perceived latency without changing the task.
- Creative drafting guardrails: explicitly distinguish source-backed facts from creative wording. Tell the model "do not invent specific names, first-party data, metrics, roadmap status, or customer outcomes; if no citable support, write a generic draft with placeholders." Prevents the slide-deck/launch-copy hallucination class.
- Verification loops for high-impact actions: name the validation commands. For coding agents, "run targeted unit tests, type checks, build checks, or a minimal smoke test" rather than vague "verify your work."
- For 5.4: reasoning effort is a last-mile knob, not a default-up dial. Higher effort is not always better, choose by task shape.
- For 5.4 manual replay of assistant messages: preserve the `phase` field exactly (`commentary` for intermediate, `final_answer` for completed). Skipping this is the top cause of "model treats intermediate update as final answer" bug.
- Codex skill at `github.com/openai/skills/tree/main/skills/.curated/openai-docs` will auto-migrate a project's prompts via `$openai-docs migrate this project to gpt-5.5`. Worth running on any 5.x stack older than 6 months.