I spent most of my first year using ChatGPT typing things like “write me a blog post about email marketing” and then wondering why the output read like every other blog post on the internet.
The fix was not a better model. The fix was the prompt.
After working through OpenAI’s prompt engineering guide, Anthropic’s Claude prompting best practices, and three peer-reviewed papers on prompting techniques (Wei et al. 2022, Kojima et al. 2022, Yao et al. 2022), I landed on seven prompt patterns that actually move the needle. They work on ChatGPT, Claude, and Gemini because they target how large language models behave, not quirks of one product.
If you are still typing one-line prompts and hoping for magic, this guide walks through what to stop doing, what to start doing, and why each pattern works.
“Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, the model will be too.” - Anthropic, Prompting Best Practices
Why Most ChatGPT Prompts Fail in 2026
Most ineffective prompts share the same problem. They are vague about what success looks like.
OpenAI’s prompt engineering team puts it bluntly: a “lack of specificity in the prompt is a common cause of underperformance” (OpenAI Prompt Engineering Guide). Vague instructions leave the model guessing. Guessing produces generic output. Generic output is what you already saw last week on the same topic.
The second failure mode is treating reasoning like magic. The 2022 Kojima paper showed that simply appending “Let’s think step by step” lifted GPT-3’s accuracy on the GSM8K math benchmark from 10.4% to 40.7% with zero other changes (Kojima et al., 2022). The hidden lesson: telling the model to think before it answers is a real lever, not folklore.
A third failure mode is over-prompting. Anthropic’s docs explicitly warn that “instructions like ‘CRITICAL: You MUST use this tool when…’ will cause overtriggering” on newer models because they are “more responsive to the system prompt than previous models” (Anthropic Prompting Best Practices). On a 2026-vintage model, polite specificity beats shouting.
The seven patterns below work because they fix one of those three failure modes each.
What “Ineffective” Actually Looks Like
Here is what ineffective ChatGPT prompts tend to share. I see these patterns over and over when people send me their “broken” prompts.
- No role or context. “Write a landing page” leaves the model inventing every variable.
- No success criteria. The model has no way to know if it has finished well.
- Negative framing only. “Don’t be boring, don’t use jargon, don’t sound like AI” tells the model what to avoid, not what to do. Anthropic’s guidance is explicit: “Tell the model what to do instead of what not to do.”
- All instruction, zero example. Especially for tone or format, the model has nothing to anchor on.
- Dumping everything at once. A 1,200-word prompt trying to write a novel, a brand guideline, a content calendar, and an SEO audit at the same time.
If your prompt does three or more of those things, no model upgrade will save you.
The 7 Prompt Patterns That Work
Each pattern below includes the research backing, a copy-pasteable template, and the one trap to avoid.
1. Role + Context (Set the Person, Explain the Why)
Tell the model who it is acting as, and why it should care. Then it can choose tone, depth, and trade-offs without you specifying each one.
A bare role line is cheap. Anthropic’s docs show the difference between "You are a helpful coding assistant specializing in Python" and a prompt that just says "How do I sort a list?" (Anthropic, System Prompt). Even a single role sentence produces noticeably tighter answers.
Template:
You are a [role] working for [audience]. You care about [one to three values].
Task: [what you want].
Constraints:
- [Length, format, or channel]
- [One thing to do or avoid]
- [How you will judge the answer]Why it works. The role primes vocabulary and assumptions. The constraints tell the model what “done well” means.
Common trap. Naming five roles at once (“act like a journalist, marketer, and SEO expert”) dilutes all of them. Pick one. Add context instead.
2. Chain-of-Thought / Step-by-Step
For anything involving math, multi-hop reasoning, or careful analysis, force the model to think out loud before answering. Wei et al. showed this in the original Chain-of-Thought paper with GSM8K math word problems (Wei et al., 2022). Kojima et al. showed the same effect with zero examples needed, just the magic phrase “Let’s think step by step.”
The lift is not subtle. MultiArith accuracy went from 17.7% to 78.7% with that single appended line in the original 2022 study.
Template:
[Your task or question]
Before you answer, walk through your reasoning step by step.
At the end, give me the answer in [format].Why it works. Models do better when they commit to intermediate steps before emitting a final answer. The reasoning trace acts as scratchpad.
Common trap. Asking reasoning models like GPT-5.6 to “think step by step” can backfire. OpenAI’s reasoning best practices are clear: “Avoid chain-of-thought prompts” for o-series because they already reason internally (OpenAI, Reasoning Best Practices). Use CoT prompting when reasoning is off or on a non-reasoning model.
3. Few-Shot / Exemplar Pattern (Show, Don’t Tell)
For tone, format, or structure, examples beat descriptions. Every time. Anthropic recommends 3 to 5 examples per prompt and suggests wrapping them in <example> tags so the model can tell instructions from data (Anthropic, Prompting Best Practices).
Template:
Task: [what you want]
Format each item as:
- Item: [short label]
- Why it matters: [one sentence]
- Fix: [concrete suggestion]
Example 1:
- Item: Passive voice in the headline
- Why it matters: Headlines need to punch, not whisper
- Fix: "We doubled signups" instead of "Signups were doubled by us"
Example 2:
[Another one in the same shape]
Now do the same for this draft:
[Your text]Why it works. Few-shot examples anchor format and tone more precisely than any verbal description. They also let you show the level of detail you expect without listing it.
Common trap. Examples that contradict the rules you wrote above them. The model averages them. Make sure your examples follow your own spec.
4. Structured Output / Schema Pattern
If the output needs to land in a spreadsheet, database, JSON object, or downstream tool, ask for a schema. OpenAI’s Structured Outputs feature guarantees the model will produce JSON that conforms to a JSON schema you supply, eliminating “missing key” or “wrong enum” errors (OpenAI, Structured Outputs).
You don’t need a developer setup to use the pattern. Even in chat, asking for a table or a fixed list of fields forces the model to organize its thinking.
Template:
Extract the following fields from the text below:
- name (string)
- role (one of: founder, marketer, engineer, other)
- pain_points (array of strings, max 3)
- next_action (one of: call, email, none)
Return the fields in a markdown table with headers.
Text:
"""
[paste]
"""Why it works. Structure forces the model to commit to specific slots, which raises accuracy on extraction tasks. Anthropic also notes that long context prompts should put queries at the end so the relevant context stays anchored (Anthropic, Prompting Best Practices). Their internal tests showed queries-at-the-end improved response quality by up to 30% on multi-document inputs.
Common trap. Asking for structure plus “and also explain your reasoning.” Pick one. Mixed output is hard to parse downstream.
5. Constraint + Success Criteria Pattern
Tell the model what success looks like in measurable terms. Length, audience, format, or a list of must-have elements. This is what Anthropic’s docs call “be very specific about your end goal” when prompting reasoning models (OpenAI, Reasoning Best Practices).
Template:
Write a [deliverable] for [audience].
Success criteria:
- Word count between [X] and [Y]
- Mentions [1, 2, 3] by name
- Avoids [banned term] entirely
- Ends with a single concrete call to action
- Tone is [reference: e.g., "like a senior PM writing to her team"]
If you cannot meet all criteria, tell me which one you dropped and why.Why it works. Without success criteria, the model optimizes for “plausible.” With them, it optimizes for a target. The “tell me which one you dropped” line is a small but powerful trick borrowed from Anthropic’s code review guidance: it forces the model to be honest about trade-offs instead of silently breaking rules (Anthropic, Prompting Claude Opus 4.8).
Common trap. Listing 15 criteria. Three to five is the sweet spot. Past that, the model starts ignoring some.
6. Multi-Step / Chained Prompt Pattern
For complex work, do not ask for everything in one shot. Break the task and let each step feed the next. The ReAct paper from 2022 was one of the first to formalize this: interleave reasoning, action, and observation in sequence, and the model can solve problems it cannot solve in a single forward pass (Yao et al., 2022).
You do not need a research agent to apply this. Even in chat, chaining works:
- Step 1: “Summarize the source text in five bullet points.”
- Step 2: “Turn those bullets into three LinkedIn post drafts.”
- Step 3: “Critique each draft for tone and clarity. Pick one and refine it.”
Why it works. Chaining isolates each decision so the model gives each one its full attention. It also gives you a checkpoint to redirect if step 1 goes sideways.
Common trap. Chaining too much. Three to four steps is plenty. If you need ten, you probably need an agent or a workflow tool.
7. “Do This, Not That” / Positive Framing Pattern
This is the one Anthropic calls out by name in their style guidance. “Tell the model what to do instead of what not to do.” Their example: instead of “Do not use markdown,” say “Your response should be composed of smoothly flowing prose paragraphs” (Anthropic, Communication Style).
Research backs this intuition. A 2023 Stanford NLP paper, DSPy, found that pipelines with declarative positive constraints routinely beat negative-prompt pipelines by 25% or more, because negative instructions are easy to skip and hard to honor (Khattab et al., 2023).
Template:
[Original ask]
Voice: warm, direct, written by a senior practitioner who has shipped this.
Reading level: eighth grade.
Format: three short paragraphs, no bullet lists.
Opening: a one-sentence hook that names the reader's real problem.
Closing: one question for the reader.Why it works. Positive framing gives the model a destination instead of a list of walls. It also tends to produce output that feels less robotic.
Common trap. Writing the negative version first and then mechanically flipping it. Test your flipped version. Some negatives simply do not have a clean positive form. In those cases, keep the negative but add context, as Anthropic suggests: explain why a constraint matters (e.g., “your response will be read by a text-to-speech engine, so never use ellipses since the engine will not pronounce them”).
Comparison: Which Pattern to Use When
Patterns are not equal. Here is how they stack up against the common tasks I see people struggle with.
| Pattern | Best for | Strength | Watch out for | Cost (tokens) |
|---|---|---|---|---|
| Role + Context | Almost any general task | Cheap and high-impact | Vague roles (“creative”) | Low |
| Chain-of-Thought | Math, logic, multi-hop analysis | Big lift on reasoning | Backfires on o-series / GPT-5.6 reasoning mode | Medium (longer output) |
| Few-Shot Exemplars | Tone, format, brand voice | Anchor format precisely | Conflicting examples | High (3–5 examples) |
| Structured Output | Extraction, downstream tools, tables | Reliable, parseable | Mixed prose and structure | Medium |
| Constraint + Success Criteria | Anything with a target | Forces measurable outcomes | Too many criteria | Low |
| Multi-Step Chaining | Research, content workflows | Better on complex tasks | Too many steps slow you down | High across turns |
| Positive Framing / “Do This” | Voice, style, polish | Avoids robotic output | Not every negative has a clean positive | Low |
My default for most tasks: Role + Context first. If the output is sloppy, add Success Criteria. If it is off-brand or off-format, switch to Few-Shot Exemplars. Reserve Chain-of-Thought for reasoning tasks. Use Structured Output the moment you copy-paste into a spreadsheet or a script.
Why These Patterns Transfer Across Models
In June 2026, OpenAI is rolling back reusable prompt objects from the API and steering developers to “treat prompts as application code,” storing them in version control with tests (OpenAI, Prompting Guide). The reasoning is that the patterns are durable. The specific phrasing you tune today will probably outlast the model it was tuned against.
That tracks with what I see in practice. The same role-plus-context structure works on GPT-5.6, Claude Opus 4.8, and Gemini 2.5. The same few-shot pattern works in the API and in the consumer ChatGPT app. The model is the easy variable to change. The prompt is the hard one to maintain.
How to Build Your Own Stacked Prompt (Example)
Here is how I would combine three of the patterns to write a useful landing page intro for an imaginary product.
You are a senior product marketer writing for founders and PMs who
are skeptical of "AI for X" claims. You care about clarity, evidence,
and respect for the reader's time.
Task: Write the first 120 words of a landing page for [Product], which
helps [audience] do [outcome] without [common pain].
Success criteria:
- Word count: 110 to 130
- Mentions the reader's real problem in the first sentence
- Names one specific mechanism the product uses (not "AI-powered")
- Ends with a concrete, falsifiable claim
- Tone: confident, plain, no exclamation points
Avoid hype words ("revolutionize," "10x," "cutting-edge").
Draft three options, each in its own <option> tag. Briefly note which
one you'd publish and why (one sentence).This single prompt stacks Role + Context, Success Criteria, Few-Shot (via the <option> tag format), and Positive Framing. The result is three options that you can pick from in 90 seconds instead of regenerating until you burn out.
A Quick Checklist Before You Hit Send
Before you run any prompt, run it through this checklist.
- Does it name the role and audience in one sentence each?
- Are success criteria measurable (word count, list of fields, format)?
- Are examples provided for format or tone?
- Are positive instructions used wherever the constraint has a clean positive form?
- Have I avoided telling a reasoning model to “think step by step”?
- Is anything that should be a follow-up prompt kept out of this one?
If two or more answers are no, rewrite before you send. Better prompts in, better prompts out.
Limitations and Honest Caveats
I want to be straight about a few things.
The 30% number is internal. Anthropic notes that “queries at the end can improve response quality by up to 30% in tests, especially with complex, multi-document inputs.” That is a directional number from their own labs, not a peer-reviewed universal claim. For simple prompts it will be much smaller; for very long context it can matter more.
Few-shot helps but does not replace instruction. The Anthropic guide cautions that examples that contradict your instructions will be averaged. So few-shot is a layer on top of clarity, not a substitute.
Reasoning models changed the rules. The “think step by step” trick still works on small chat models but, per OpenAI’s reasoning best practices, can actually degrade performance on the o-series. If you are on GPT-5.6 in reasoning mode, prefer direct goal statements plus constraints over CoT phrasing.
I do not have a number for “7.” Seven is a usable number for this article. The underlying patterns are well supported by research and the official guides. Whether it is five or nine patterns you keep is less important than applying the ones that match your work.
A Pull Quote You Can Steal
A prompt is not a wish. It is a specification.
If you only internalize one thing, let it be that. Specificity is the difference between a model that guesses and a model that performs.
Sources
These are the resources I leaned on for this guide.
- OpenAI. Prompt Engineering Guide. platform.openai.com/docs/guides/prompt-engineering
- OpenAI. Prompting. platform.openai.com/docs/guides/prompting
- OpenAI. Reasoning Best Practices. platform.openai.com/docs/guides/reasoning-best-practices
- OpenAI. Structured Model Outputs. platform.openai.com/docs/guides/structured-outputs
- OpenAI. Text Generation. platform.openai.com/docs/guides/text
- OpenAI. Model Spec (February 2025). model-spec.openai.com/2025-02-12.html
- Anthropic. Prompt Engineering Overview. docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
- Anthropic. Prompting Best Practices. docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/claude-prompting-best-practices
- Anthropic. Prompting Claude Opus 4.8. docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/prompting-claude-opus-4-8
- Anthropic. Extended Thinking. docs.anthropic.com/en/docs/build-with-claude/extended-thinking
- Wei, Jason, et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” arXiv:2201.11903 (2022). arxiv.org/abs/2201.11903
- Kojima, Takeshi, et al. “Large Language Models are Zero-Shot Reasoners.” arXiv:2205.11916 (2022). arxiv.org/abs/2205.11916
- Yao, Shunyu, et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” arXiv:2210.03629 (2022). arxiv.org/abs/2210.03629
- Khattab, Omar, et al. “DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines.” arXiv:2310.03714 (2023). arxiv.org/abs/2310.03714
- OpenAI et al. “GPT-4 Technical Report.” arXiv:2303.08774 (2023). arxiv.org/abs/2303.08774
Wrapping Up
If you only have time to change one thing today, change how you close your prompts. End them with a clear success criterion in measurable terms, a named audience, and a specific format. That single rewrite will outperform almost any new model upgrade in your workflow.
The seven patterns stack on top of each other. Pick two or three that match your most common task. Build a reusable template. Save the prompt itself somewhere you can version it. As OpenAI is now saying out loud, prompts are code. Treat them that way and they will pay you back for years.