TL;DR
Most “bad” ChatGPT output is a missing-instruction problem, not a model problem. Add four things to any weak prompt — audience, format, constraint, example — and you go from roughly 1-in-5 usable answers to 4-in-5. Then drop the habits that GPT-5.5 made obsolete: “think step by step,” 1,000-word context dumps, and rigid “ALWAYS/NEVER” rules. Since GPT-5.5 became the ChatGPT default on April 23, 2026, OpenAI’s own guidance is shorter, outcome-first prompts that say what good looks like and let the model find the path.
Why your prompts feel mid
Daily users hit a plateau: every response feels average because they keep sending the same 30-word ask and hoping. The fix is rarely a fancier “magic prompt.” It is closing the gap between what you pictured and what you actually typed.
This guide is the 10-minute upgrade. It is for anyone using ChatGPT daily — writers, PMs, marketers, students, developers — who can’t tell whether a weak answer is the model’s fault, the prompt’s fault, or the task’s fault. Reach for it when:
- You keep tweaking the same response without real improvement.
- You get fluent-but-wrong output: generic, off-tone, or technically incorrect.
- You can’t articulate why a “good” answer would be good. That blank is the actual problem.
The four levers (add in this order)
- Audience — who reads this, and what do they already know? “Explain to a CTO who already runs Kubernetes” produces radically different output from “Explain to a marketer.” Audience sets depth and vocabulary in one line.
- Format — bullets, table, short paragraph, JSON, headline plus three sentences. Left alone, GPT-5.5 defaults to plain prose (OpenAI now recommends prose as the model’s default for explanations), which is almost never the shape you need.
- Constraint — length cap, tone, words to avoid, what NOT to do. “Under 120 words” and “no marketing adjectives” are the two highest-leverage constraints most people skip.
- Example — paste one paragraph of “good” output, yours or someone else’s. A single concrete sample beats two paragraphs of abstract style description every time.
Before and after
WEAK PROMPT
Write a launch email for our new feature.
UPGRADED PROMPT
Write a launch email for our new feature.
Audience: existing users who already pay us $20/mo.
Format: subject line + 3 short paragraphs + 1 CTA button label.
Constraint: under 150 words, no exclamation marks, no "excited to announce".
Example tone (one paragraph from our last launch):
"We pushed a small change today: ..."The upgraded version routinely produces something you can ship after a five-minute edit. Notice it is barely longer — it is more specific, not more verbose.
What GPT-5.5 changed (and what to stop doing)
GPT-5.5 has been the ChatGPT default since April 23, 2026, and the model picker shows Auto / Instant / Thinking / Pro. The reasoning models think before answering by default, which quietly retired three habits from the GPT-4 era. OpenAI’s official prompt guidance now says, in plain terms, that old prompt stacks are holding the model back. Drop these:
| Old habit (pre-2026) | Why it now hurts | Do this instead |
|---|---|---|
| ”Think step by step” on every prompt | GPT-5.5 reasons by default; the phrase adds noise | Just state the goal; ask for reasoning only on genuinely hard, multi-step problems |
| 1,000+ word kitchen-sink prompts | Anticipating every edge case narrows the model’s search space | Start with the smallest prompt that captures the contract, then add only what failed |
| ”ALWAYS / NEVER / must / only” everywhere | Locks judgment calls and cuts off better answers | Use decision rules: “Ask a clarifying question only if missing info would change the answer" |
| "You are a world-class expert…” persona | Vague personas do little on a model already tuned to be capable | Specify behavior, not status: “Respond the way the author of Refactoring would” |
Save absolutes for true invariants — safety rules, required output fields, things that must never happen. For everything else (“when should it search the web?”, “when should it ask a follow-up?”), an “if X, then Y, otherwise Z” rule beats a blanket command. This is OpenAI’s own recommendation for GPT-5.5, not folklore.
Where to put each instruction
You have three places to put guidance, and mixing them up is a common source of mid output.
| Where | What belongs there | Persistence |
|---|---|---|
| Custom Instructions (Settings) | Your standing role, tone, and “how I like answers” defaults | Every new chat |
| Project Instructions (a ChatGPT Project) | Per-project rules, reference files, recurring constraints | Every chat inside that Project |
| The chat prompt | This specific ask: the four levers for the task in front of you | This conversation only |
Rule of thumb: anything you’d repeat in three different chats belongs in Custom or Project Instructions, not pasted in again. For Custom GPTs, keep the system prompt tight — concise instructions (well under ~8,000 characters) outperform sprawling ones.
The loop: one variable at a time
- Write your first-draft prompt as you normally would.
- Read it back: “Could 50 different people with 50 different audiences have written this?” If yes, you’re missing audience.
- Add audience, format, constraint, example — in that order. Stop at “good enough,” not “perfect.”
- Run it. If the output is wrong, change one variable and re-run. Changing three things at once teaches you nothing about which one mattered.
Try it once on real work: take a task you ran last week with mid output, save the original prompt and answer, rewrite with all four levers, run, and diff the two responses. If the new prompt also works on a similar task a day later, lock it in as a template.
Quality check before you ship
- Does the output answer your actual question, or a more flattering version of it? Models love to restate the prompt charitably.
- Are the facts independently verifiable, or confident-sounding mush? GPT-5.5 still fabricates citations.
- Would your reader actually use this, or just nod politely?
Reuse: build a small prompt library
- Keep a
prompts/folder in your notes — one file per recurring task, with the four levers written in a header comment so you can edit confidently. - For team work, ship the prompt file alongside the deliverable. New hires get up to speed faster.
- Re-test templates each quarter. Model updates change defaults; a prompt tuned for an older model can quietly underperform on GPT-5.5. (OpenAI literally advises starting migrations from a “fresh baseline” rather than reusing old prompt stacks.)
FAQ
- Does a longer prompt help? Up to a point. For most chat tasks, ~150–500 words of clear instruction is plenty. Past ~1,000 words you mostly add noise and narrow the model’s options — GPT-5.5 is more sensitive to bloat than older models, not less.
- Should I use Custom Instructions or Project Instructions? Both, for anything recurring. Put role and tone in Custom Instructions; put per-project rules and reference files in a Project. Keep per-chat prompts focused on the single task.
- Do “tip the model $200” or “my career depends on this” tricks work? They were always marginal, and on GPT-5.5 they’re noise. Skip them and spend the effort on a concrete example instead.
- Will newer models make prompt skill obsolete? Better models reduce some prompt sensitivity, but audience and format will always matter — the model can’t read your mind. Notably, GPT-5.5 changed which habits help, so the skill shifts rather than disappears.
- What changed most with GPT-5.5 specifically? Three things: it reasons by default (so “think step by step” is redundant), it prefers outcome-first over step-by-step process prompts, and it responds better to decision rules than to ALWAYS/NEVER commands.
For the official source, see OpenAI’s prompt engineering guide and GPT-5.5 model guidance.