The first prompt should be a single block that names the role, the reader, the angle, the length, the format, one worked example, and one hard constraint — and you paste it before you ask the AI to write a single sentence.
Here is the template, ready to copy: "You are a [role, e.g. freelance finance writer]. Write for [specific reader, e.g. first-time landlords in the UK]. The angle is [one sentence, e.g. 'why a £200 boiler service saves money']. Target [length, e.g. 1,200 words] as [format, e.g. an H2-and-bullet explainer]. Include one worked example: [e.g. a two-bed flat in Leeds where a £90 service avoided a £600 repair]. Constraint: do not mention [the thing you want excluded, e.g. commercial mortgages]. Do not write yet — reply with an outline of five H2 headings and the word count for each."
The last line is the part most people skip, and it's the part that decides whether the draft is usable. Asking for an outline first forces the model to commit to a structure you can correct in ten seconds, before it has spent a thousand words going the wrong way. If the outline is wrong, you rewrite one line. If the draft is wrong, you rewrite everything.
Why the template is built in that order
Each slot in the template closes off a different way the output can miss. The role sets vocabulary and assumed expertise — a "freelance finance writer" writes differently from a "friendly general blogger," even on the same topic. The reader sets what gets explained versus what gets assumed, which is where most beginner drafts fail: they explain compound interest to an accountant, or skip it for a first-time saver.
The angle is the single most important slot, and it should be one sentence you could argue with. "Why a £200 boiler service saves money" is an angle. "Boiler maintenance" is a topic. A topic has thousands of possible articles inside it; an angle has one.
The worked example is the slot people leave out, and it is the one that stops the model drifting into abstraction. Give it real inputs — a place, a number, a person — and the rest of the draft tends to stay concrete because it has something to anchor to. The constraint is a negative instruction, and it works better than a positive one. "Do not mention commercial mortgages" is checkable. "Keep it focused" is not.
A decision rule for how many constraints to use
A useful rule: include exactly one hard constraint, and make it the one that prevents the failure you have actually seen before. Not the failure you imagine — the one that happened last time. If your last post on this topic drifted into tax law when you wanted practical maintenance advice, your constraint is "do not discuss tax treatment." If it kept recommending tools you don't cover, your constraint is "do not name software products."
Three or more constraints in a first prompt tends to backfire. The model starts negotiating with itself about which rules matter, and you get a draft that obeys the small rules and ignores the big one. One constraint, checked after the outline, beats five constraints checked after the draft.
Where this template breaks down
The template assumes you already know your angle and your reader. If you don't, the first prompt should be a research prompt instead — ask the model to list five possible angles for the topic with a one-line argument for each, then pick one and run the template above. That's a two-step process, and it's slower, but it beats templating an angle you haven't chosen.
It also assumes the model has enough context to write accurately about your subject. For niche or fast-moving topics, no prompt fixes a knowledge gap. You'll get confident, well-structured prose about a version of reality that doesn't exist. The fix is to supply the facts yourself in the prompt — paste three or four sentences of source material and tell the model to use only those. That's slower to set up and it's the honest trade-off: a longer prompt buys you less editing later.
On tooling, the slot to use is the system or instruction field, not the chat box. In ChatGPT that's Custom Instructions or a Project; in Claude it's a Project's instructions; in Gemini it's a Gem. Putting the role, reader, and constraint there means you only type the angle, length, and example each time.
That's the difference between a prompt you write once and a prompt you retype every morning. According to our AI tool database, which tracks 360 AI tools with pricing and capability snapshots verified as recently as 2026-09-18, the tools differ far more in how they store standing instructions than in the quality of a single well-scoped prompt — so pick the one whose instruction field you'll actually use.
One last thing worth knowing: the template works because it front-loads the decisions you'd otherwise make while editing. It doesn't make the model smarter. It makes your corrections cheaper, which is the only lever you really have.