A Prompt Structure That Survives Model Upgrades

A prompt survives a model upgrade when you separate what you want from how one model likes to be asked, and when you keep a small test set that tells you whether the new model still delivers. Anthropic, OpenAI and Google all document this same split in their own prompting guides, which I read on 13 November 2026. This post turns it into a six-part spec you can reuse.

Key Takeaways

• Anthropic's prompt engineering overview assumes you already have success criteria, a way to test against them and a first draft. Write the criteria before the prompt.
• Anthropic's best practices page says any technique tied to a named model was measured on that model and should be re-checked against your own evals before you reuse it elsewhere.
• OpenAI's prompt engineering guide recommends sections in this order: identity, instructions, examples, context. It also says to pin production apps to model snapshots and to build evals that catch regressions on upgrade.
• Parameters and settings are the fragile part. Anthropic lists temperature, top_p and top_k as removed in API version 2026-10-21, returning a 400 error. Google's Gemini docs warn that newer models ignore those values and that future releases will return errors.
• A six-part spec (goal, inputs, constraints, format, examples, acceptance checks) is portable. Vendor syntax and sampling settings belong in a thin adapter layer you can swap.
• Ten to twenty real test cases, scored the same way before and after an upgrade, beat any amount of prompt cleverness.

Why prompts break on upgrade

Prompts rarely break because the new model is worse. They break because the prompt contained two kinds of content mixed together: the task definition, and workarounds for a specific model's habits. When the model changes, the workarounds stop helping and sometimes start hurting.

Anthropic's migration notes show the pattern. They say Claude 4.6 models are more proactive and may overtrigger on instructions that older models needed, so earlier "be thorough" and "use tools aggressively" language should be reduced. The same page says prefilled responses on the last assistant turn are unsupported from Claude 4.6 onward, and that manual thinking budgets return a 400 error from Claude 4.7 on. Those are all model-era details that sat inside someone's prompt template.

The practical rule: anything that would still make sense if you handed the prompt to a new human contractor belongs in the spec. Anything that only makes sense for one model belongs in an adapter you can delete.

The six-part prompt spec

Google's Gemini prompting guide lists clear instructions, constraints, response format, few-shot examples and context as core strategies. OpenAI and Anthropic list nearly the same elements under different names. The spec below merges them.

1. Goal and audience. One sentence on the job and who reads the result. Anthropic's guide describes the model as a capable new employee without your context, so state the context.
2. Inputs. Name each input and wrap it in its own labelled block (XML tags or Markdown headers; both vendors accept either, and consistency matters more than the choice). Anthropic's page says labelled tags reduce misreading when instructions, context and examples are mixed.
3. Constraints. Must-do and never-do lists, including what to do when information is missing. For factual work, include a permission to say "not in the source." Anthropic's hallucination guidance says explicitly allowing "I don't know" can reduce false information.
4. Output format. Table, headings, word limits, fields. Say exactly what you want.
5. Examples. Anthropic suggests 3 to 5 examples that are relevant, diverse and wrapped in <example> tags. Gemini's guide warns that too many can cause overfitting to their pattern, so keep them varied and few.
6. Acceptance checks. A short list the output must satisfy, such as "every claim has a source from the input" or "no sentence over 25 words." You will reuse these as your scoring sheet.

OpenAI's guide places context near the end of the prompt, and Google's says to put instructions at the end for long contexts. Vendors differ on placement, so treat placement as an adapter choice and test it.

Before and after: a hypothetical example

The following is an illustrative example I wrote for this post. The brand and course are invented.

Before (typical, vague):

Write a landing page section about our data analytics bootcamp. Make it
engaging and SEO friendly. Be thorough. Don't make things up.

Problems: no audience, no source material, "engaging" and "thorough" mean different things to different models, "don't make things up" gives no fallback, and nothing can be scored.

After (spec form):

GOAL: Draft a 120-150 word "Who this is for" section for a landing page.
AUDIENCE: Working professionals in operations roles considering a career
switch into analytics. They have not coded before.

<inputs>
<program_facts> ...paste the approved fact sheet here... </program_facts>
<brand_voice> ...three approved sentences from the live site... </brand_voice>
</inputs>

CONSTRAINTS:
- Use only facts in <program_facts>. If a needed fact is missing, write
  [MISSING: fact name] instead of guessing.
- No outcome promises (jobs, salary, timelines) unless they appear in
  <program_facts> word for word.
- Second person, plain sentences, no sentence over 25 words.

FORMAT: One heading, one paragraph, then a 4-item bulleted list.

<examples> ...3 short approved sections for other programs... </examples>

ACCEPTANCE: Before answering, check each constraint and list any
[MISSING] items after the draft under "Gaps".

Nothing in the second prompt depends on which vendor you use. If you later need an adapter, it might add a system-message wrapper for one API, or an effort setting for another. The spec stays the same.

The adapter layer

Keep adapter settings in one place: a config file, a variable block or a note at the top of the prompt document. Typical contents:

• Model name or snapshot pinned (OpenAI recommends pinning snapshots so behavior stays consistent).
• Sampling parameters, if the platform still accepts them. As of 13 November 2026, Anthropic's deprecations page says they were removed in API version 2026-10-21 and recommends prompting instead, while Gemini 4 and 3.6+ ignore custom values per Google's guide.
• Thinking or effort settings. Anthropic's notes say Claude 4.6 and later use adaptive thinking and an effort parameter, and manual budgets error on 4.7 and later.
• Prompt-caching layout, if you use it. See prompt caching for marketers for when it is worth organizing around.

If you work only in a chat interface, your adapter is almost empty. The test set is what protects you.

The 10-case regression set

Anthropic's success criteria guide says good criteria are specific, measurable, achievable and relevant, and that evals should mirror the real task and include edge cases. It also says more questions with automated grading beat fewer hand-graded ones. For a marketing team without engineering support, a scaled-down version works.

1. Pull 10 real past inputs (briefs, fact sheets, transcripts). Include 2 that are messy and 2 where key information is missing.
2. Write the acceptance checks as yes/no questions.
3. Run the current prompt and record pass or fail per check.
4. When a new model arrives, run the same 10 inputs with the same prompt, change nothing, and record results.
5. Only then edit the prompt, one change at a time, re-running the set.
6. Log the date, model, score and what you changed.

My full model-comparison protocol is in how to evaluate AI models for marketing. If several people share prompts, the ownership and versioning ideas in governing a team prompt library sit on top of this.

What this means for marketers

• This week: pick your single most-used prompt and rewrite it into the six parts. Do not change what it is meant to do.
• Write the acceptance checks first. If you cannot state a check as yes or no, the prompt has no way to be wrong, which means you cannot improve it.
• Remove push-harder language. Lines like "be thorough" or "use tools aggressively" were workarounds for older models. Anthropic's migration notes say to reduce them for Claude 4.6 and later because those models may overtrigger. My own guess is that capital-letter emphasis ages the same way, so re-test without it.
• Keep three examples, not ten. Vary them on purpose.
• Leave alone: prompts that pass their tests. Upgrading the model does not oblige you to rewrite the prompt.
• Treat vendor tips as dated. My read is that any tip with a model name in it expires faster than the structure around it.

For ready-made starting prompts by task, see Claude prompts for marketing teams, and fold them into this spec.

Common mistakes

• Writing a long prompt before defining what a good result looks like.
• Testing a new model on one favorite prompt and one lucky output.
• Mixing model workarounds into the task description so nobody knows which lines still matter.
• Not logging changes, which makes a regression impossible to trace.
• Leaving "don't make things up" as the only factual safeguard instead of supplying sources and a "missing" fallback.

Caveats

The vendor docs I used are living pages, and parameters and tips will change again, which is the point of the adapter layer. I have not run these specific prompts against any model, and the example above is illustrative, not a measured result. The numbers 10 and 3 are my working defaults, not vendor figures.

FAQ

What is the best prompt structure for marketing work?

A six-part spec works across vendors: goal and audience, labelled inputs, constraints with a "missing information" fallback, output format, a few varied examples, and acceptance checks. OpenAI, Anthropic and Google each describe close variants of these elements in their own guides.

Do I need to rewrite prompts every time a model updates?

Not by default. Run your saved test cases on the new model first. If they still pass, leave the prompt alone. If some fail, change one element at a time and re-run.

Should I use XML tags or Markdown headers in prompts?

Both are accepted. Anthropic recommends consistent descriptive XML tags, and OpenAI suggests Markdown headers and XML tags to separate sections. Pick one style, apply it consistently and test whether a different style changes results for your task.

How many examples should a prompt include?

Anthropic suggests 3 to 5 relevant and diverse examples. Google's guide cautions that too many can lead the model to overfit to them. Start with three that differ from each other.

Why did my prompt stop working after a model change?

Usually because it relied on a model-specific habit, such as strong emphasis or manual thinking settings, that the new model treats differently. Anthropic's migration notes describe this for Claude 4.6 and later. Check the adapter layer first, then the examples.

Are sampling settings like temperature still worth tuning?

It depends on the platform. As of 13 November 2026, Anthropic documents their removal in API version 2026-10-21 and Google documents that newer Gemini models ignore them. Check your vendor's page before relying on them.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help turning your team's prompts into tested, reusable specs, see my work and reach me through the contact form at younusfardeen.com.