Chaining and Decomposition - Big Tasks, Small Prompts
10 min read
Ask an AI to "create my whole marketing plan" in one prompt and you will get something that looks impressive and helps almost nobody — shallow everywhere, deep nowhere. Big tasks fail in single prompts for the same reason they fail when you delegate them in one breath to a person: too many decisions crammed into one instruction.
The professional move is decomposition — breaking a big job into steps — and chaining — feeding the output of each step into the next prompt. You already work this way; now your prompts will too.
Why one giant prompt fails
When a single prompt asks for research, strategy, and finished copy all at once, the model must average its attention across everything. Worse, you never get a checkpoint. If step one heads the wrong direction, every step built on it inherits the error — and you only find out at the end. Chaining gives you a steering wheel between every step.
A real chain: the quarterly promotion
Leo owns a small gym and wants a member-referral campaign. Instead of one mega-prompt, he runs a chain:
Prompt 1 (Explore):
I run a 300-member gym in a mid-size city. Members are mostly
25-45, value community over equipment. Suggest 5 referral
campaign ideas with a one-line pitch and the main risk of each.Leo picks idea number 3 and steers:
Prompt 2 (Develop):
Let's develop idea 3, the "bring a friend for free week." Draft
the campaign outline: the offer, the rules (keep them simple),
the timeline over 4 weeks, and how staff should mention it at
the front desk.He fixes one rule he does not like, then continues:
Prompt 3 (Produce):
Using the outline above, write the announcement email to members.
150 words, energetic but not pushy, one clear call to action.Prompt 4 (Adapt):
Now adapt that email into: a) an Instagram caption with 3
hashtags, b) a 40-word script for front-desk staff.Four small prompts, a human decision between each. Total time: fifteen minutes, and every piece is on-strategy because Leo steered at the checkpoints.
How to decompose any big task
- Draft the steps first. Ask yourself: if I gave this to a capable assistant, what would they hand me at each check-in? Those check-ins are your prompts.
- One deliverable per prompt. If the output would need two different headings, it should probably be two prompts.
- Put your judgment between steps, not after all of them. Pick, cut, and correct at each stage — that is where your expertise does the most work.
- Carry forward only what matters. In a long chain, start a fresh chat at a natural break and paste in just the approved outputs so far, not the whole meandering history.
A handy pattern to remember: Explore, then Develop, then Produce, then Adapt. Generate options, deepen the chosen one, create the main asset, then spin off variations. It fits marketing campaigns, hiring processes, event planning, new-service launches — almost any multi-part project on your plate.
You can even ask the AI to plan the chain for you: "I want to launch a catering side-business from my food truck. Break this into 6 steps I could work through with you one prompt at a time." Then run the steps together.
Try it now
Pick a genuinely big task you have been putting off — a seasonal campaign, an employee handbook, a new service launch. First ask the AI to propose a step-by-step chain for it. Edit the steps so they match how you actually work, then run just the first two prompts today. Notice how different the quality feels when you steer between steps instead of grading one giant answer at the end.