How to update prompts for GPT-5.5 and modern AI models

Executive overview

Detailed, step-by-step mega-prompts are becoming counterproductive with today's frontier models. These models can determine the best path to a goal on their own — over-directing them limits their intelligence.

The replacement is a four-part structure: specify the destination, define what good looks like, build in doubt (proof requirements), and set a finish line (done).

The smartest use of a modern AI model is to tell it where to go, not how to get there.

Destination: specify intent, not steps

• Old approach: tell the AI the exact steps ("do step 1, step 2, step 3...")
• New approach: state the outcome and the reason behind it
• Bad: "Summarise this meeting transcript"
• Better: "Turn this transcript into a follow-up email I can send to a client"
• Bad: "Make a table from this spreadsheet"
• Better: "Find the three problems in this spreadsheet that would change my decision for X"
• The AI determines the most effective path; you state only where you're headed

Definition: binary success criteria

• Specify what good looks like in your specific context — brand voice, format, verifiable claims
• Prefer binary criteria over spectrum-based ones — the AI can check its own work against yes/no conditions
• Example: "Keep it under 200 words" and "Put the ask in the first three sentences" are binary
• Vague criteria ("be clear and calm") are harder to self-audit; binary criteria get closer to correct on the first pass

Doubt: require proof for every claim

• Frontier models are right more often but also guess more confidently — a dangerous combination
• For any output involving financial, legal, or brand decisions, require inline citations
• Bad: "Don't make stuff up" or "Don't hallucinate" — these instructions have little effect
• Better: "After every fact or claim, cite the source inline — include the report and page number"
• Better: "When you're not sure, write 'unverified' or leave it blank — I'd rather see a gap than a guess"
• The second instruction changes the AI's incentive: a blank answer is acceptable; a fabricated one is not

Done: set a finish line

• High-reasoning modes (e.g. "extra high" in Codex, "heavy" in ChatGPT) can run for hours — wasting time and tokens
• Most tasks don't need exhaustive reasoning; set an explicit stopping condition
• Bad: "Be exhaustive", "cover every angle", "think deeply"
• Better: "Stop once you can answer the main question with enough evidence"
• Better: "When the output meets the checklist, give me the final version"
• Tie the finish line to the specific task and any checklist you've provided

Putting the four Ds together

Old mega-prompt structure (to retire):

• "Act as a world-class strategist. First read the transcript. Then identify themes. Then extract action items. Then write the email..."

New four-D structure (same task):

• Destination: "Turn this transcript into a client-ready follow-up email"
• Definition: "Success means the email clearly states what we decided, what is still open, and the next action for each person"
• Doubt: "Use only decisions directly supported by the transcript. Put unclear items under open questions"
• Done: "When the checklist is met, give me the final email"