Prompting fundamentals that outlive model releases

You have probably collected a few phrases that seemed to work. “You are a world-class copywriter.” “Take a deep breath and think step by step.” “This is very important to my career.” They went into a notes file, and for a while the outputs got better, or seemed to. Then a new model shipped, the same phrases produced flatter results, and you started collecting again.

That collection was never the skill. The skill is much duller and much more durable: saying what the job is, to someone who has never met you. Vendors describe the same three moves in slightly different words. Anthropic says to treat the model as a brilliant but new employee who lacks context on your norms and workflows [2]. Google says clear and specific instructions are an effective and efficient way to customise model behaviour, and recommends always including a few examples [4]. OpenAI’s guide structures a prompt as identity, instructions, examples, and context [3]. Three vendors, one instruction: brief it properly. This guide is for solo operators and small teams using the chat interfaces. If you are building an application with evaluation suites and version-controlled prompts, you want the vendor docs directly, not this.

The three levers are context, constraints, and examples

Everything that reliably improves an output is one of three things you can add. Context is the material and the situation: the actual email thread, the actual invoice, who reads the result and what they will do with it. Constraints are the boundaries: length, format, what to leave out, what to do when information is missing. Examples are one or two pieces of finished work that look like what you want back.

These three survive model releases because they are not tricks played on a system’s quirks. They are information the model genuinely does not have. No release will ever make your client’s tone of voice inferable from nothing. OpenAI’s guide puts context near the end of the prompt, since you may supply different context for different requests, and uses Markdown headers and XML tags to delineate where one piece of content begins and ends [3]. Anthropic recommends wrapping examples in <example> tags so the model can tell them from instructions [2]. The formatting differs slightly between vendors; the underlying need — separate the job from the stuff — does not.

The test Anthropic gives is the one I use: show your prompt to a colleague with minimal context on the task and ask them to follow it, and if they would be confused, the model will be too [2]. It is unglamorous and it catches almost everything. Most bad prompts are not badly engineered. They are briefs that assume the reader was in the meeting.

Context means the material itself, not a description of it

The single most common failure is describing your input instead of pasting it. “Summarise the feedback we got from the workshop” gives the model nothing to summarise, so it invents plausible workshop feedback. Paste the notes and the same sentence works.

Context also includes the reason. Anthropic’s docs make this explicit with a small example: “NEVER use ellipses” is listed as less effective than explaining that the response will be read aloud by a text-to-speech engine that will not know how to pronounce them, because Claude is smart enough to generalise from the explanation [2]. The version with a reason handles the cases you did not think to list. Semicolons, stray asterisks, an em dash — all covered by “this gets read aloud”, none covered by a rule about ellipses.

So say who reads the output and what happens next. “This goes to a client who has already paid and is chasing delivery” produces different, better work than “write a professional email”. Google’s guidance is the same shape: include in the prompt the instructions and information the model needs to solve the problem, instead of assuming it already has everything required [4].

Constraints work best as a short numbered list

Write your boundaries as a list of five or six short rules, not as a paragraph of preferences. Order matters when steps depend on each other. Anthropic recommends giving instructions as sequential steps in numbered lists or bullet points when the order or completeness of the steps matters [2], and OpenAI’s guide recommends Markdown headers and lists to mark distinct sections of a prompt and communicate hierarchy to the model [3]. Either way, the shape is a list, not a paragraph.

Good constraints are checkable. “Under 150 words” is checkable. “Concise” is not. “British spelling, no bullet points, no greeting because it is a reply in a thread” is checkable. “Match our brand voice” is not, unless you attach the voice as an example. Google’s guidance on constraints is the plain version: to control length, say “summarize this text in one sentence”, and to control shape, ask for the response as a table, a bulleted list, a sentence, or a paragraph [4].

Include one constraint about missing information. Something like “if the notes do not say when the deadline is, write TBC rather than guessing”. This is the cheapest defence against invented detail there is, and it is the first technique in Anthropic’s hallucination guidance: explicitly give the model permission to admit uncertainty, which the docs say can drastically reduce false information [5].

Two examples beat two paragraphs of adjectives

If you want output that sounds like you, the fastest route is not to describe your voice. It is to paste two things you wrote and say “match this”. Anthropic calls examples one of the most reliable ways to steer output format, tone, and structure, and suggests three to five for best results, chosen to mirror your actual use case and varied enough that the model does not pick up an unintended pattern [2]. Google recommends always including a few examples and says prompts without them are likely to be less effective [4].

The diversity point is the one people miss. If both your example emails are apologies for a late delivery, you have taught the model that your voice is apologetic. Pick two that differ — one warm, one firm — and the model learns the range rather than the mood.

Build a small folder of your own examples: two emails, one proposal section, one summary you were happy with. It takes twenty minutes once and it outlasts every model in the folder’s lifetime.

Magic phrases break because they were never load-bearing

Role-play openers and incantations sit in a different category from the three levers. “You are a senior consultant” occasionally helps, because it narrows vocabulary and assumed audience, and role prompting is still listed as a technique in Anthropic’s overview [1]. But it is doing a weak version of what context does properly. “Write this for a reader who runs a two-person agency and has not heard of retainers” is a role instruction that carries actual information. “You are a world-class business expert” is flattery aimed at a system that does not have a career.

Phrases that lean on quirks are the ones that expire. Models are trained, tuned, and retired on a schedule — Anthropic’s deprecation page defines active, legacy, deprecated, and retired states, says Anthropic gives at least 60 days’ notice before retiring a publicly released model, and suggests thorough testing of your applications with the new models well before the retirement date [6]. When behaviour shifts under you, a prompt built on context, constraints, and examples degrades gracefully; you tighten one constraint. A prompt built on a phrase that used to unlock something has no repair path, because you never knew why it worked.

A practical rule: if you cannot explain what a line in your prompt tells the model that it could not otherwise know, delete the line and see whether anything changes. Usually nothing does.

Rewriting beats re-rolling

When an output is wrong, the instinct is to regenerate. Sometimes that works and you have learned nothing. The better move is to name the specific defect and change the prompt: too long means add a word count, wrong tone means add an example, invented specifics mean add the missing-information rule and paste the source.

Google’s docs frame this as iteration: different words or phrasing often yield different responses even when they mean the same thing, changing the order of the content in the prompt can change the answer, and when the model will not follow your instructions you can give it an analogous task that achieves the same result [4]. Change one thing at a time. Change three and the next output is better for reasons you cannot reuse tomorrow.

Keep the prompts that survive this process in a plain text file with a name that says when to use them. Not a system. A file. When a new model lands, run the three or four you rely on most, and fix the ones that drifted — the same testing habit Anthropic recommends before a migration [6].

What still goes wrong

Invented facts remain the main hazard, and none of this eliminates them. Anthropic’s own guidance is blunt: the techniques significantly reduce hallucinations but do not eliminate them entirely, and critical information should always be validated, especially for high-stakes decisions [5]. Invented sources are the sharper edge of the same problem, because a fabricated citation looks exactly like a real one. If a claim carries a URL, open the URL. For long documents, Anthropic suggests asking the model to extract word-for-word quotes first, then reviewing each claim against a supporting quote and removing any claim it cannot support [5].

Length drifts. You ask for 150 words, the reply comes back well over, and the overshoot is not consistent between runs of the same prompt. Nothing fully fixes this in a chat interface. Ask for the count, check it, and if length is load-bearing — a bio field with a hard limit, an ad slot — count the words yourself before it ships.

And some outputs are not yours to ship on a model’s say-so. Anthropic’s Usage Policy, effective 15 September 2025, requires that when its products are used to give advice, recommendations, or subjective decisions directly affecting individuals in high-risk areas, including legal interpretation, healthcare decisions, insurance, financial and investment advice, employment and housing determinations, and academic admissions, a qualified professional in that field reviews the content or decision before it is finalised or disseminated. Where model outputs are shown directly to the person, the policy also requires disclosing that AI helped produce the advice, at a minimum at the start of each session [7]. That is not a prompting problem and no phrase solves it. If your work touches those areas, the review step is part of the job.

Prompts from this guide

freelancer-brief

You are drafting {deliverable} for {audience}.

Context — the material to work from is below the line. Use only what is
there. The reader will use this to {reader_action}.

Rules:
1. {length_limit}
2. Format: {format}
3. Do not include {exclusions}
4. If a fact is missing from the material, write TBC rather than guessing
5. Match the tone of the example below, not a generic professional tone

Example of the tone I want:
{voice_example}

---
{material}

reply-in-my-voice

Write a reply to the message below. It is a thread reply, so no greeting
and no sign-off.

About the situation: {situation}
What I want the reader to do next: {desired_outcome}
Length: {length_limit}

Here are two emails I wrote, so you can match my voice. They are
deliberately different in tone — learn the range, not the mood.

Example 1:
{my_email_one}

Example 2:
{my_email_two}

Message to reply to:
{incoming_message}