Prompt engineering is the skill of communicating with AI systems in ways that consistently produce better outputs. The word engineering makes it sound more technical than it is. In practice it is a set of learnable habits: how to give context, how to specify constraints, how to break complex tasks into steps, how to tell the AI what you want and what you do not want. Most people learn none of these habits because nobody explains them clearly. This post does.
Everything
here has been tested directly. The examples are not illustrative. They are
actual prompts that were run and whose outputs were compared against weaker
versions of the same request. The quality difference is consistent and
significant.
How to use this guide:
Eight sections
covering the most important prompt engineering principles, each with tested
examples. Read straight through or jump to the technique most relevant to what
you are working on. The techniques are cumulative: combining multiple principles
produces better results than applying any single one.
Why Prompts Matter So Much
The same AI model produces dramatically different outputs depending on how a request is framed. This is not an accident or a flaw. It reflects how large language models actually work: they generate the most likely continuation of the input they receive. A vague input produces a vague continuation. A precise, contextualised input produces a precise, contextualised continuation.
Understanding
this changes how you approach AI tools. The question is not 'what can this AI
do?' The question is 'what information does this AI need to produce the output
I actually want?' Prompt engineering is the practice of systematically
providing that information.
Testing Note: The same task, summarise a business
proposal for a non-technical investor, was given to Claude in two versions.
Version 1 was the bare request with no additional context. Version 2 included
the investor's background, the level of technical detail appropriate, the
desired length, and two specific questions the summary should answer. Version 2
produced output that required 15 percent editing. Version 1 required 55 percent
editing to reach the same quality standard. Same model, same task, dramatically
different results.
Technique 1: Assign a Role Before Every Significant Task
Assigning
a specific role or persona to the AI at the start of a prompt consistently
produces more targeted and useful output than asking it to respond as a generic
assistant. When you tell Claude to respond as a specific type of expert
addressing a specific type of audience, it calibrates the vocabulary, the
assumed knowledge level, the structure, and the practical focus of its response
accordingly.
Weak prompt: Explain cryptocurrency to me.
Stronger prompt: You are a financial educator explaining
cryptocurrency to a 55-year-old who has only ever invested in traditional bank
savings accounts and is concerned about risk. Explain what cryptocurrency is,
why people invest in it, and the two most important risks they should
understand before considering it. Use no jargon. Keep it under 400 words.
Testing Note: The stronger prompt produced an
explanation that a non-technical test reader described as immediately
understandable and genuinely useful. The weaker prompt produced a technically
accurate but vocabulary-heavy explanation that the same reader described as
requiring prior knowledge to follow. The role assignment and audience
specification made the difference.
Technique 2: Specify Every Constraint Explicitly
AI models are highly responsive to constraints when they are stated explicitly but they do not reliably infer unstated constraints from context. If you need the output to be a specific length, specify it. If you need it to avoid specific words or phrases, list them. If you need it to include specific elements, name them. If you need it to avoid a specific tone, describe what you do not want as precisely as you describe what you do.
Claude
in particular handles multi-constraint prompts with remarkable precision. In
testing, providing six explicit constraints and asking Claude to satisfy all of
them produced outputs meeting all six in the majority of attempts. The same is
not true of models that interpret constraints loosely, which is why this
technique produces more consistent results with Claude than with some
alternatives.
Multi-constraint prompt example: Write a 200-word LinkedIn post about the benefits of remote
work. Requirements: open with a question, not a statement. Include one specific
statistic (you may note it requires verification). Do not mention productivity.
Close with a call for readers to share their experience. Write for senior
professionals aged 35-55. Do not use the words 'amazing', 'game-changing', or
'leverage'.
Testing Note: This prompt was tested on Claude and
ChatGPT. Claude satisfied all six constraints on the first attempt. ChatGPT
satisfied five, using the word 'game-changer' in a paraphrased form that
technically avoided the exact banned word. Claude's instruction-following
advantage is most visible in multi-constraint prompts with six or more
requirements.
Technique 3: Provide Context That the AI Cannot Know
AI
models have no knowledge of your specific situation, your organisation, your
audience, your previous work, or your personal preferences unless you provide
that information in the prompt. This is the most consistently underused
opportunity in everyday AI use. The more relevant context you provide, the more
specifically the output can be tailored to your actual situation rather than to
a generic version of your task.
Useful
context to include: who you are and what your role is, who the audience for the
output is, what has already been tried or decided, what constraints apply that
are not obvious from the task description, what success looks like for this
specific use case, and any relevant background the AI needs to understand why
this task matters.
Context-rich prompt structure: I am a [role] at [type of organisation]. I am preparing [type of
output] for [specific audience]. The key thing my audience needs to understand
is [core message]. The main concern they are likely to have is [objection or
concern]. Previous attempts to address this produced [describe what did not
work and why]. Please [specific task] with this context in mind.
Technique 4: Use Chain-of-Thought for Complex Reasoning
Chain-of-thought prompting is a technique backed by published research from Google Brain that significantly improves AI performance on reasoning tasks. The technique is simple: before asking for a conclusion or recommendation, ask the AI to reason through the problem step by step. The process of explicitly generating intermediate reasoning steps produces more accurate final answers than asking for the answer directly.
The
practical application is straightforward. For any task involving analysis,
decision-making, or problem-solving, add the instruction 'reason through this
step by step before giving your final answer' or 'think through the key
considerations before making a recommendation.' The quality improvement on
complex analytical tasks is consistent and significant.
Chain-of-thought prompt:
I am deciding
whether to expand my catering business to a second location. Before giving me a
recommendation, reason through the following step by step: what information
would a business analyst need to assess this decision, what are the strongest
arguments for expanding now, what are the strongest arguments against, what is
the single most important factor I should resolve before deciding, and then
give me your overall assessment.
Testing Note: The same expansion decision question was
asked with and without chain-of-thought instruction. Without it, both Claude
and ChatGPT produced a balanced but shallow list of pros and cons. With the
chain-of-thought instruction, both produced structured reasoning that
identified a key missing information gap (lease terms and capital availability)
that the direct-answer version had not surfaced. The chain-of-thought version
was substantially more useful for actual decision-making.
Technique 5: Show the AI What Good Looks Like
Few-shot prompting is one of the most powerful and most underused techniques in everyday AI use. Rather than describing what you want, you show the AI an example of it. Provide one, two, or three examples of the type of output you want before giving the actual task, and the model will pattern-match against those examples to produce an output consistent with their style, format, and quality.
This
is particularly valuable for tasks with a specific house style that is
difficult to describe in words but easy to demonstrate. A social media manager
whose posts have a distinctive voice, a legal professional whose documents
follow a specific structural convention, a teacher whose explanations use a
particular scaffolding method: all of these benefit from showing the AI what
the desired output looks like rather than attempting to describe it.
Few-shot prompt structure: Here are two examples of the type of [output] I want: [Example
1]. [Example 2]. Please produce [number] more examples in exactly the same
style and format for the following [content/topic]: [your actual task].
Testing Note: A brand voice that had been difficult to
describe verbally was demonstrated through two example social media posts.
Claude produced four additional posts that the brand owner rated as matching
the voice more closely than the five posts produced without examples in the
previous session. The few-shot approach is consistently more effective than
voice descriptions alone when the target style is distinctive but hard to
specify in abstract terms.
Technique 6: Use Negative Constraints to Eliminate Unwanted Patterns
Telling the AI what not to produce is as important as telling it what to produce. AI models have default tendencies: toward formal language, toward bullet-pointed structure, toward diplomatic balanced conclusions, toward certain transitional phrases that signal AI generation to experienced readers. Explicitly prohibiting these defaults breaks the pattern and produces more distinctive output.
Negative
constraints that consistently improve output quality include: do not use bullet
points, write in continuous prose; do not use the word 'however' as a sentence
opener; do not reach a balanced conclusion, take a clear position; do not
summarise what you are about to say before saying it; do not end with a generic
call to action; do not use the passive voice. Each of these directly addresses
a default AI writing pattern.
Negative constraint example: Write a 300-word opinion piece arguing that remote work is
better for productivity than office work. Constraints: take a clear,
unambiguous position without hedging. Do not use bullet points. Do not use the
phrases 'it depends', 'on the other hand', or 'there are many perspectives'. Do
not summarise the argument in the opening sentence. Open with a specific
scenario or observation, not a thesis statement.
Technique 7: Iterate Within the Conversation, Not Across Conversations
One of the most common mistakes in AI use is starting a new conversation when an output is unsatisfactory instead of refining it within the existing conversation. This wastes the context the AI has already built up about your task, your preferences, and your previous feedback. Every refinement instruction you give in a continuing conversation makes the AI better calibrated to your specific needs for that session.
Effective
refinement instructions are specific about what changed and what stayed the
same. Rather than 'try again', say 'the structure is good but the opening is
too formal. Rewrite just the opening paragraph with more conversational energy,
keeping everything else identical.' Rather than 'make it shorter', say 'this is
400 words. Cut it to 250 words. Preserve the second and fourth paragraphs
unchanged and condense the first, third, and fifth.'
Effective refinement prompt: The overall structure and argument are good. Three specific
changes: the opening sentence is too generic, replace it with something more
specific and attention-grabbing. The third paragraph is too long, cut it by
half while keeping its main point. The closing line is weak, replace it with
something that leaves the reader with a specific takeaway rather than a vague
invitation.
Technique 8: Ask the AI to Identify What It Does Not Know
AI models do not spontaneously acknowledge uncertainty with the frequency or specificity that would be most useful. They are trained on feedback that rewards confident, complete-sounding responses, which creates pressure to produce fluent outputs even when the model is uncertain about specific claims. Asking the AI directly what it is most uncertain about in its own response counteracts this tendency and produces more honest, more useful answers.
The
technique: after receiving an analytical response, ask 'On a scale of one to ten,
how confident are you in each of the main claims in your last response, and
where should I verify independently before relying on this?' The resulting
answer consistently identifies the specific claims that deserve verification,
which is more useful than a general disclaimer.
Calibration prompt:
Looking back
at your previous response, rate your confidence in each of the three main
claims from one to ten. For any claim you rate below eight, identify
specifically what would need to be verified and where I should look to verify
it.
Testing Note: When this calibration prompt was applied
after a research synthesis on a technical policy topic, Claude correctly
identified the two claims that were later found to require minor factual
correction when verified against primary sources. ChatGPT's calibration
response was less specific, identifying general areas of uncertainty rather
than specific claims. Both were more useful than outputs that included no
uncertainty acknowledgment at all.
Putting It All Together: The Full Prompt Framework
The eight techniques above work individually and compound when combined. For any significant task, the full framework looks like this: assign a role and audience, provide all relevant context, specify every constraint including negative ones, show an example if you have one that represents the desired output, ask for step-by-step reasoning if the task involves analysis, treat the first response as a draft and refine within the conversation, and ask for a calibration check on any factual claims that matter.
This
sounds like a lot of work for a simple request. For simple requests it is. The
framework is for tasks where the output quality genuinely matters: professional
documents, important communications, complex analysis, content that will
represent you publicly. For quick lookups, casual questions, and low-stakes
tasks, a short prompt is fine. The discipline is knowing which category your
task is in.
The AI Vanguard Take:
Prompt
engineering is not a technical skill. It is a communication skill applied to a
new medium. The people who are getting the best results from AI tools are not
the people who know the most about how transformers work. They are the people
who have learned to be precise, contextual, and iterative in how they
communicate what they want. That is a skill anyone can develop, and developing
it will compound across every AI tool you use for the rest of your working
life.
Frequently Asked Questions
Does
prompt engineering work differently on different AI tools?
The core principles apply across all major AI tools. Claude is particularly responsive to multi-constraint prompts and role assignments. ChatGPT responds well to few-shot examples and chain-of-thought instructions. Gemini benefits most from context-rich prompts that give it enough information to leverage its web access meaningfully. The calibration technique works best with Claude, which tends to give more specific uncertainty assessments than the alternatives.
Will AI
tools eventually require less prompting as they improve?
Models are getting better at inferring intent from less explicit prompts, and this trend will continue. However, the gap between a well-crafted and a poorly crafted prompt is likely to persist because better models amplify the quality advantage of better inputs rather than eliminating it. The skill of communicating clearly and precisely with AI systems is becoming more valuable, not less, as the models become more capable.
Are
there any AI tools that help you write better prompts?
Yes.
Claude and ChatGPT can both help you improve your own prompts. Describe what
you are trying to achieve and ask the AI to help you write a more effective
prompt for that task. This meta-prompting approach is genuinely useful for
learning, because the AI will often surface context and constraints that you
had not thought to include. You can also ask any AI tool to critique a prompt
you have already written and suggest specific improvements.
The Deep Dives Category:
The AI Vanguard
publishes deep-dive posts on how AI works, AI ethics, and the mechanics of AI
tools every week. Subscribe below to receive every post in your inbox.
