Beyond the prompt recipe: iterative collaboration, not a better template

Overview

The recipe everyone learns has four parts — persona, task, context, format. It’s a fine way to get unstuck on your very first prompt. It’s also the wrong model for how good work with AI actually happens now.

Why now: even the vendor that teaches the recipe — Google’s Gemini for Workspace guide — tells you to drop the parts you don’t need and to treat prompting as a conversation. The frontier has moved further still: Anthropic’s prompting guidance for Claude Fable 5 (June 2026) tells users to have the model scope the task and ask clarifying questions before executing. The skill is no longer composing the perfect instruction up front; it’s working the task out with the model.

What you’ll be able to do: spot when to stop filling a template and start collaborating — and use two moves that beat any better-worded prompt.

The content

The recipe is seductive because it looks like a card you fill in: name a persona, state the task, add some context, specify a format, and out comes the answer. For a genuine first-timer it does help — naming the slots surfaces the context you’d have left out. Google’s own guidance names the four, then says in the same breath that you “don’t need to use all four” — and its companion prompt guide reports that the most fruitful prompts average around 21 words with relevant context, while the prompts people actually try are usually under nine.

Here’s the overturn. The parts don’t stop mattering; they relocate. The strongest practitioners rarely assemble a template up front — they discover the task with the model and curate the context as they go. “Prompting” quietly becomes collaboration, and two moves do most of the work.

Start generic, then iterate. Point the model at the task in a line or two and let the details surface over the next few messages, rather than cramming them into one long brief where they get buried. A reliable opener is to make the model interview you: “Before you draft anything, ask me the questions you’d need answered to do this well.” You converge on the real task together — which is usually not the task you’d have written down cold.

Let the model write the next prompt. Carrying a project from one session to the next, don’t re-explain it from memory — ask the outgoing session to write the prompt that starts the new one, then edit it. The model packs the context it actually used; you keep what matters and cut what doesn’t. Those hand-off prompts come out sharper and more task-specific than anything you’d draft from a cold start.

So treat the four-part recipe as training wheels, not the bike. One caveat: collaboration isn’t a licence to ramble — for a two-line task, a single clear instruction still wins. The shift is for work that’s bigger than one ask, which is most real work.

Try it

Take a task you’d normally write a careful prompt for. Don’t. Open with one line and this:

I want help with: [your task in one line].

Before you do anything, ask me the 5–7 questions you'd most need
answered to do this really well — about the goal, the audience, the
constraints, and what "done" looks like. Wait for my answers before drafting.

Answer its questions, then let it proceed. Notice how much of what surfaces you’d never have thought to put in a one-shot prompt. Where this won’t help: trivial, self-contained asks — if a single sentence fully specifies the job, just ask it.

Additional reading

• Writing effective prompts — Gemini for Google Workspace (Google) — names the four parts and says you “don’t need to use all four.” The recipe, from the people who teach it.
• Gemini for Workspace prompting guide (PDF) (October 2024) — source of the length finding: the most fruitful prompts average ~21 words with relevant context; people usually try fewer than nine.
• Prompt engineering overview — Anthropic (living page) — frames prompting as empirical and iterative — set success criteria, test, refine — and defers to the living best-practices reference.
• Prompting Claude Fable 5 — Claude Docs (June 2026) — vendor doctrine for the newest frontier model: have it scope the task and ask clarifying questions before executing, and carry context across sessions deliberately.
• Effective context engineering for AI agents — Anthropic (2025-09-29) — why curating context beats composing a template; the principle behind a good hand-off prompt.
• Tobi Lütke on “context engineering” (X) (2025-06-19) — state a problem with enough context that it is “plausibly solvable” — exactly what a model-written hand-off prompt is for.

Editor’s note

My own prompting has dropped the recipe almost entirely. When I carry a project between sessions, the outgoing session writes the next session’s prompt and I amend it — the context lands more completely than anything I would reconstruct by hand. For new work I start deliberately generic and let the real task surface over a series of short messages, rather than burying details in one long brief. The shift worth making is from prompting to iterative collaboration.