Prompts That Compose

I stopped writing prompts as one long sentence and started building them out of parts. The day I did, editing a prompt stopped feeling like rewriting it.

In the first lesson I said to direct Claude like a collaborator rather than summon it like an oracle. This lesson is the mechanics of how. When I brief a person well, I don't say everything in one breath — I give them their role, the task, the limits, an example of what good looks like, and the form I want the result in. A good prompt has exactly those same parts. The word I use is compose: build the prompt out of pieces, the way you'd compose anything from components.

Why bother naming the parts? Because the moment a prompt is a structure instead of a paragraph, you can change one part without touching the rest. The answer came back too long? Edit the output shape line — leave everything else alone. It missed the audience? Fix the role. A wall-of-text prompt has no seams, so every edit means rewriting the whole thing and hoping you didn't break what was working. A composed prompt has seams on purpose.

A worked example

Take a request I make often — turning rough notes into a short summary. The lazy version is the one line above. The composed version names the parts, and you can watch each one earn its place:

# role        You are summarising for a busy colleague who missed the meeting.
# task        Turn the notes below into something they can read in 30 seconds.
# constraints Three bullets. Decisions only — skip the discussion.
# example     "- Shipping Friday, not Monday (scope cut)"
# output      Return only the three bullets. No preamble.

Nothing here is clever. But every line removes a way the answer could go wrong — too long, too vague, the wrong format, an unwanted “Here is your summary!” on top. And tomorrow, when I want six bullets instead of three, I change one word and re-run. That's the payoff of seams.

A few moves that punch above their weight

Inside those parts, a handful of habits reliably change the result. Give the why behind a constraint — “no ellipses, because a screen reader reads them aloud” — and the model generalizes from the reason instead of obeying narrowly. Say what to do, not what to avoid: “write in flowing prose” beats “don't use bullet points.” Ask it to act, not suggest: “rewrite this paragraph” gets a rewrite; “could you suggest changes” gets a list you still have to apply. And one that's changed: you no longer need to shout. Older advice said to pile on “CRITICAL!!! you MUST” — current models follow plain instructions and actually over-react to the loud ones. Write it like you'd write it to a person.

Composing across prompts

The second meaning of compose is chaining. A task too big for one good prompt usually breaks into a short pipeline, where each prompt does one thing and hands its output to the next. Asking for “a finished blog post” in one shot gives you mush; asking for an outline, then drafting each section from that outline, then a final pass for tone — three prompts — gives you something you'd actually publish. Each prompt's output is the next one's input.

Try it yourself

Take a prompt you reuse and split it into labelled lines — role, task, constraints, example, output. Run it once. Then change exactly one line and run it again. Feeling how a single-line edit reliably moves a single thing about the answer is the whole skill; once you have it, you stop rewriting and start adjusting.

Grounded in Anthropic's prompt-engineering guidance and its writing on chaining prompts.