CARE says every strong prompt needs Context, an Anchor, restrictions and an example, which is one answer to what makes a good prompt; leave out any one of the four and the quality drops noticeably.
- Context stops generic output. Tell the model the audience, purpose and background, even when it feels obvious.
- Anchors carry meaning. A named style, camera or publication locks the output to something concrete.
- Restrictions block defaults. Say what you do not want, such as filler phrases or plastic skin.
- Examples work best. Showing an example is the most reliable way to get consistent output.
- It works across models. CARE describes what a prompt must contain, not how to phrase it for one model.
CARE describes what a prompt must contain, not how to phrase it for a particular model. That is why it has not dated while phrasing tricks have — the reasoning is on context engineering.
CARE is a four-part prompt structure that works across every type of AI model. It's not a formula you fill in mechanically — it's a checklist of what every effective prompt contains. Once you know it, you'll use it automatically.
The framework is especially useful for LLM prompting, where structure matters more than it does for image generation. But the underlying logic applies anywhere: give the model context, anchor it to a specific output, restrict what it shouldn't do, and show it an example of what you want.
Context is everything the model needs to know about the situation before it starts generating. Who is the audience? What's the purpose? What's the background? What does the model need to understand to give a useful response?
Most people skip context because it feels obvious to them. It isn't obvious to the model. Without context, the model defaults to the most generic interpretation of your request. With it, the model can tailor its output to your specific situation.
For LLMs: "I'm writing a product description for a £400 moisturiser targeting women 35–50 who buy from Net-a-Porter. They're informed buyers who find over-explained ingredient lists patronising."Not measured: an invented example brief written for this page (2026); the price is part of the scenario, not market data.
For image prompts: The environment, time of day, and setting are context. "Rain-soaked Tokyo street at midnight, neon reflections on wet pavement" sets the context for every visual decision that follows.
An anchor is a specific reference point that locks the model's interpretation to something concrete. It can be a named style, a real camera, a publication, a format, or anything else that carries specific meaning the model has been trained on.
A filmmaker’s name is the riskiest anchor. "Shot in the style of Roger Deakins" is usually shorthand for natural practical sources, restrained colour and deep shadow — the quality of light in No Country for Old Men — but what a model takes from a name is unpredictable. Write those words into the prompt instead of the name; what a director’s name does to a prompt covers what carries across and what does not.Attribution checked 11 Sep 2026: Roger Deakins was the cinematographer on No Country for Old Men (2007, dir. Joel and Ethan Coen) — Wikipedia and Rotten Tomatoes' interview with Deakins about the film. The description of what that anchor evokes is this page's reading, not a sourced claim about how any model behaves.
For LLMs: "Write in the tone of a Kinfolk magazine feature" or "Format as a two-column comparison table." Both are anchors — they invoke a specific, recognisable output the model has encountered in training.
For image prompts: The camera and lens are the most reliable anchor. "Canon EOS R5, 85mm f/1.2" anchors the output to the depth, compression, and quality associated with that specific combination.
Restrictions are the things you don't want. They're easy to overlook because they feel like negatives — you're describing absence, not presence. But they're often the most important part of a prompt because they eliminate the model's most predictable failure modes.
Without restrictions, models default to their highest-probability outputs. For image models that means plastic-looking skin, flat lighting, stock photo composition. For LLMs it means corporate filler phrases, hollow enthusiasm, vague generalisations.
For LLMs: "No filler phrases. Don't start sentences with 'I'. No exclamation marks. Avoid the words 'curate', 'elevate', 'transform'. Under 100 words."
For image models: The negative prompt is the restriction field. CGI, plastic skin, airbrushed, overexposed, watermark, blurry — put your most common failure modes here every time.
Showing the model an example of what you want is the single most reliable way to get consistent output. Examples communicate things that descriptions can't. A model that's shown a reference image, a sample paragraph, or a worked example will reproduce the pattern far more reliably than one given only an abstract description.
For LLMs: a few-shot prompt includes examples of input-output pairs before your actual request. The model reads the pattern and applies it. One example helps (“Here’s a product description that’s close to what I want. Write something similar for this product:”); for a reliable pattern, Anthropic suggests three to five.Anthropic, Prompting best practices, as quoted on the LLM cheat sheet: “Include 3–5 examples for best results.” Corrected 22 Sep 2026: this said one example “significantly improves results”, with no source.
For image models: reference images via ControlNet or IP-Adapter — two different things, not one — or conversational iteration in ChatGPT Images 2.5 are the equivalent. Showing is usually more precise than describing.Corrected 11 Sep 2026: this line read "ControlNet IP-Adapter", as though it were a single tool. It is two. Read at source 11 Sep 2026 — IP-Adapter project page and its paper: "IP-Adapter is an effective and lightweight adapter to achieve image prompt capability for pretrained text-to-image diffusion models," whose "key design … is a decoupled cross-attention mechanism that separates cross-attention layers for text features and image features," and which "does not change the original network structure" so it can be combined with ControlNet. ControlNet is the separate conditioning network. ChatGPT Images 2.0 was verified on 11 Sep 2026 as OpenAI's name for the image model (gpt-image-2), launched 21 Apr 2026 — openai.com, listed in search but not read at source; the name and date were confirmed in contemporaneous coverage. The sentence also previously said showing is "always" more precise than describing; that absolute is not sourced, so it now reads "usually". It had already been superseded: OpenAI released ChatGPT Images 2.5 on 8 Sep 2026 (openai.com, read in a browser 16 Sep 2026; OpenAI blocks automated fetches), and this page now names 2.5.
Even if you can't provide a literal example, you can describe one: "The tone should feel like the copy on Aesop's product pages — dry, specific, no superlatives." That's a reference the model can use.
CARE in Practice — Full Example
Not measured: an invented example brief written for this page (2026); the £280 price and the shopper are part of the scenario, not market data.
That prompt will produce output that's usable on the first pass. Remove any one of the four elements and the quality drops noticeably.
All four parts together give usable output on the first pass. The example sets the buyer, names a tone anchor, bans certain words, caps the length and shows a sample.
More prompts, when something changes.
Prompts, model guides and workflow notes, sent when there is something new. Free.
SUBSCRIBE FREE ↗