Constraints and iteration
Set limits that shape the answer, let the assistant ask you questions before it starts, and improve a weak answer with specific feedback instead of starting over.
LIMITS HELP
Constraints narrow the guess
Every limit you state removes thousands of possible answers you didn't want. The most useful ones are length ("under 80 words"), scope ("only the costs, not the history"), must include ("mention the refund deadline") and avoid ("no bullet points, no English terms"). Without them the model falls back on its defaults: fairly long, fairly formal, a bit of everything.
"Summarise this article" gives you 300 words. "Summarise this article in 3 sentences for someone deciding whether to read it" gives you what you needed.
Check yourself
Negar asks for "a short bio for my design portfolio" and gets 180 words in the third person with phrases like "passionate creative". Which set of constraints fixes the most?
- "Make it much, much shorter and way less cheesy, and make it sound like a real person talking about their own work, not a brochure."
- "Make it better and more modern."
- "Shorter please."
- "Under 60 words, first person, plain words, no 'passionate' or 'creative'. Must mention: 6 years in packaging design, Shiraz."
Show the answer
"Under 60 words, first person, plain words, no 'passionate' or 'creative'. Must mention: 6 years in packaging design, Shiraz."
Yes. Length, voice, banned words and must-haves, each one checkable. She can look at the result and see at once whether it obeyed.
FLIP THE INTERVIEW
Let it ask you first
Sometimes you don't know what context matters. Hand that problem to the model: "Before you start, ask me the questions you need answered to do this well." It knows what a good answer to this kind of task depends on, and its questions pull the missing context out of your head. Cap the number, or you may get fifteen.
Useful for anything personal or open-ended: a study plan, a difficult message, a trip, a business idea.
I want a 3-month plan to get my English ready for a job interview. Before you write anything, ask me up to 5 questions you need answered to make the plan fit me. Ask them all at once, then wait for my answers.Happy to. Five questions first:
1. What is your level now? (For example: can you follow a film without subtitles?)
2. What job is the interview for, and will the whole interview be in English?
3. How many minutes a day can you really give this, on a normal day?
4. Which is weakest: speaking, listening, vocabulary or confidence?
5. Do you have anyone to practise speaking with?Every one of these would have been a silent guess. "Wait for my answers" matters: without it, some assistants ask the questions and then write the plan anyway.
Check yourself
When the first answer isn't right, the most effective habit is to open a new chat and type the same request again, hoping for a better draw.
Show the answer
False
A new draw from the same prompt gives you a different answer with the same problem, because the missing information is still missing. Stay in the chat and say what was wrong. The conversation so far is context you've already paid for.
STEERING
Refine, don't restart
Good feedback has three parts: what to keep, what to change, and why. The "why" matters most, because it lets the model fix the cause and not just the one sentence you pointed at. Treat it like briefing a capable assistant who has never met your boss.
"Keep the opening line. The second paragraph sounds like an advert, and my boss distrusts sales language, so make it factual. Cut the last sentence."
The refinement loop
- Read it and name the problem
Not "I don't like it". Is it too long, too formal, missing a point, wrong focus, a wrong fact?
- Give keep / change / why
One or two changes per round. Ten at once and some get dropped.
- Ask for what you need back
"Rewrite only the second paragraph" or "give me the full text again with the changes". Say which.
- Know when to restart
If the chat has become long and tangled and old mistakes keep returning, open a new one. Start it with an improved prompt that includes everything you learned along the way.
Check yourself
Which of these follow-up messages will actually move the answer forward?
- "Make it better."
- "Keep paragraph 1, halve paragraph 2, drop the joke."
- "No, that's wrong."
- "Too formal. Write it like a message to a colleague I like."
- "Try again."
- "Option 2 is closest. Give me 3 more like it, but shorter."
Show the answer
Specific feedback: "Keep paragraph 1, halve paragraph 2, drop the joke.", "Too formal. Write it like a message to a colleague I like.", "Option 2 is closest. Give me 3 more like it, but shorter."
Too vague to help: "Make it better.", "No, that's wrong.", "Try again."
Check yourself
Ali is drafting a message to decline a wedding invitation. The draft is warm but gives a long, awkward excuse. What should he send next?
- "This isn't good. Write another one."
- Open a new chat and type the original request again, this time adding that it should be short and warm, and hope for a better draft
- "Keep the warm opening and the good wishes. Cut the excuse: I just want to say I can't come, no reason. 3 sentences at most."
- "I don't like it, can you fix it?"
Show the answer
"Keep the warm opening and the good wishes. Cut the excuse: I just want to say I can't come, no reason. 3 sentences at most."
Right: what to keep, what to change, why, and a limit. The next draft has almost no way to miss.
Lesson recap
- Constraints you could check with a ruler (length, scope, must include, avoid) narrow the answer fast.
- Pair every "don't" with a "do" when you can.
- "Ask me up to 5 questions first, then wait" pulls out the context you didn't know mattered.
- Refine with keep / change / why. Restart only when the chat is tangled, and carry what you learned into the new prompt.