zmime.com
General assistant · ChatGPT

Prompt for ChatGPT

Most ChatGPT problems are input problems. Set up your instructions once, then keep each request to a single job.

  • Assistant
  • Custom instructions
  • Structured output
  • Reasoning
01 — Overview

What ChatGPT rewards

ChatGPT is the default general-purpose tool for most people, which means most of its bad output comes from prompts that skip context. It will happily answer a question it had no basis to answer. The fix is unglamorous: state the role, paste the material, name the constraints and define the output.

Two habits help more than any trick. Put your standing preferences in custom instructions so you stop retyping them, and treat a chat as an iterative session — draft, critique, revise — rather than expecting one prompt to produce a finished artefact. OpenAI’s prompt engineering guide is the reference for how instructions, roles and reasoning settings interact on the API side.

  • 01Broad general capability across writing, analysis, code and planning.
  • 02Custom instructions and projects for reusable context.
  • 03Good at iterative refinement when you critique its output.
  • 04Reasoning models for multi-step problems where you can trade time for accuracy.
02 — Anatomy

A prompt structure that works for everyday tasks

  • Role + audience Whose expertise, and who the output is for.
  • Task One deliverable. Split anything with an "and then" into two turns.
  • Context The material, pasted. Plus what you already tried.
  • Constraints Length, format, tone, exclusions, reading level.
  • Output spec Headings, table, JSON, numbered list — name it.
  • Quality bar What a great answer includes, so it can self-check.
  • Uncertainty rule "Mark anything you are unsure about" beats hoping it stays honest.
03 — The prompts

5 prompts written for ChatGPT

Paste, then replace the [bracketed] parts. Keep the wording identical for anything you want repeated across a set.

Nothing matches that word. Clear the field to see every prompt.

Prompt 01

Custom instructions worth keeping

Paste into settings once; it improves every conversation afterwards.

  • setup
  • custom instructions
Template · plain text
HOW I WORK
I am a [role] working mainly on [domains]. Assume [level] technical knowledge; skip beginner explanations unless I ask.

HOW TO RESPOND
- Lead with the answer, then the reasoning. No preamble, no restating my question.
- Be concise by default. Expand only when I ask for depth.
- Plain text prose for explanations; code blocks only for code; tables for comparisons.
- No filler openers ("Great question"), no closing offers to help further.
- If I am wrong about something, say so directly and explain why.

ACCURACY
- Separate what you know from what you are inferring.
- Mark uncertain claims with [unverified] rather than smoothing over them.
- If a task needs information you do not have, ask for it instead of guessing.
- Never invent statistics, citations, prices or API methods.

DEFAULTS
- Language: [English variant]. Units: [metric/imperial]. Dates: [format].
- Code: [language and version], [style preferences], include error handling, no unexplained magic numbers.
Prompt 02

Analysis with an explicit reasoning path

Decisions and trade-offs where you need to see the logic.

  • analysis
  • decision
Template · plain text
Help me decide: [decision, stated as a question].

SITUATION: [context, in enough detail that a stranger could follow]
OPTIONS: [A], [B], [C]
CONSTRAINTS: budget [x], timeline [y], team [z], cannot change [w]
WHAT MATTERS MOST: [rank your criteria]
WHAT I HAVE ALREADY RULED OUT AND WHY: [...]

DELIVER
1. A comparison table: option against each of my criteria, with a short justification per cell.
2. The strongest case for each option, stated fairly.
3. Your recommendation, with the two assumptions it depends on most.
4. What evidence would change the recommendation.
5. What could go wrong in the first 30 days, and the cheapest way to find out early.

RULES
- No hedging conclusions. Commit, then state the risk.
- Mark anything you inferred rather than took from my brief.
Prompt 03

Critique-then-revise loop

Getting genuinely better output instead of a polite second draft.

  • editing
  • iteration
Template · plain text
Here is a draft of [artefact]. Do not rewrite it yet.

STEP 1 — CRITIQUE
Review it as [a skeptical target reader / a senior reviewer]. List the 5 weakest points, most serious first. For each: what is wrong, why it matters to the reader, and what evidence or change would fix it. Include anything that is unclear, unsupported, or padding.

STEP 2 — WAIT
Stop after the critique. I will tell you which points to act on.

CONTEXT
- Audience: [who]
- Purpose: [what it must achieve]
- Constraints: [length, tone, format]

DRAFT:
"""
[paste draft]
"""
Prompt 04

Learning a topic without the fluff

Getting up to speed on something new, fast.

  • learning
  • explainer
Template · plain text
Teach me [topic]. I already understand [related things I know], so build on that.

STRUCTURE
1. The core idea in 3 sentences, no analogies yet.
2. One analogy, and where the analogy breaks down.
3. The 5 concepts I must understand, each in 2-3 sentences, in dependency order.
4. A worked example with real numbers or real code.
5. The three mistakes beginners make and what causes them.
6. How to tell whether I have understood it: 5 questions, hardest last, with answers hidden below.
7. What to learn next, and what to deliberately ignore for now.

RULES
- Define jargon inline the first time.
- Do not simplify to the point of being wrong; flag where you are simplifying.
- Skip history unless it explains why the thing works the way it does.
Prompt 05

Data cleanup and reformatting

Turning messy pasted content into a clean, usable structure.

  • data
  • formatting
Template · plain text
Convert the content below into [a Markdown table / CSV / JSON] with these columns: [col1], [col2], [col3].

RULES
- Preserve values exactly; do not correct spellings or reformat numbers except as instructed below.
- Normalise dates to YYYY-MM-DD. Normalise currency to [format].
- Use "" for genuinely missing values. Never guess.
- If a row cannot be parsed, list it separately under "Unparsed" with the reason.
- Deduplicate only on exact matches of [key column], and tell me what you removed.
- No commentary before or after the output.

CONTENT:
"""
[paste messy data]
"""
04 — Craft notes

What separates a good prompt from a wasted run

Do this 05

  • Move standing preferences into custom instructions.
  • Keep one deliverable per turn and iterate.
  • Ask for a critique before a rewrite.
  • Require [unverified] tags on shaky claims.
  • Paste the source material rather than describing it.

Not this 05

  • "Act as an expert" with no other context.
  • Chaining five tasks into one prompt.
  • Asking for citations and accepting them without checking.
  • Continuing a long thread after it starts contradicting itself — restart with a summary.
  • Trusting arithmetic and dates without verification.
05 — Questions

Asked often, answered plainly

Do "act as an expert" prompts help?

A role helps set vocabulary and depth, but only slightly. What actually changes output quality is the context you supply and the constraints you set. A role plus no material is still a guess.

Why does quality drop in long conversations?

Earlier turns compete with later ones and contradictions accumulate. When a thread starts drifting, ask for a summary of the current state and requirements, then paste that into a fresh conversation.

When is a reasoning model worth the wait?

Multi-step problems where a wrong intermediate step ruins the answer: debugging, planning, maths, anything with constraints to satisfy simultaneously. For drafting and rewriting, a faster model is usually the better trade.

06 — Sources

Read the primary documentation

Vendors change flags, limits and defaults often. Where this site disagrees with official docs, the docs win.