Prompt decomposition

One prompt, many steps

A request like “research three vendors, compare them, flag the risks, write a recommendation and make a five-slide briefing” is really five or six jobs. Prompt Tornado reads the prompt, splits it into steps, works out which steps depend on which, picks a model for each, and shows you the plan before anything runs.

What the planner produces

A task type for each step, such as research, writing, code, image or voice. Prompt Tornado recognises 181 specific task types (an executive summary, a competitive analysis, a code review…), grouped into a small set of routing types.
An instruction written for that step alone.
Its inputs: which earlier steps it needs. Independent steps don't wait for each other's results.
A ranked list of models that can do it. If the first choice fails, the next one runs.

A real plan: the Vendor Decision Briefing

One of the four example workflows in the app, with the model that ran each step.

#	Step	Uses	Model that ran it
1	Research the three vendors from official docs, with sources	the prompt	Perplexity, then GPT-5.5
2	Turn the research into a structured comparison (JSON)	step 1	GPT-5.5
3	Review research and comparison against the requirements	steps 1–2	GPT-5.6 Sol
4	Write a decision memo recommending one vendor	steps 1–3	Claude Opus 5
5	Turn the memo into a five-slide executive outline	steps 2–4	Claude Opus 5
6	Create an illustration for the briefing	the prompt	FLUX (fal.ai)

From a production run on 16 September 2026 (total model cost $0.58). The prompt, the full plan and the unedited output are on GitHub.

You can see and change the plan first

Preview Plan shows every step, the model planned for it and the estimated cost before the run starts. You can rename a step, rewrite its instruction, or remove it. The run executes the plan you approved, not a hidden one.

How task types are recognised

Clear requests are matched to a task type by rules, so the same wording gets the same type. Where a clause is ambiguous, a language model classifies it, set to its most consistent output (temperature 0). Each of the 181 task types has a test case that checks it is routed to the expected model.

How this differs from building the workflow yourself

	You draw or code the steps	Prompt Tornado
Automation tools (Zapier, Make, n8n)	You build the flow step by step	The plan is generated from the prompt
Frameworks (LangChain)	You write the orchestration code	Managed, with routing and fallback
Model choice	Picked by hand for each step	Picked per step from ranked routes

Each step is a bounded operation: usually one model call, though a research step pairs a search model with a model that writes up the results, and a failed call falls back to the next model in line. It is not an autonomous agent loop that decides its own next action. That makes a run predictable, priced up front and reviewable; it also means open-ended tasks that need many rounds of tool use are a better fit for an agent framework.

Frequently asked

Can I edit the plan before it runs?

Yes. Preview Plan shows each step, its planned model and the estimated cost, and you can rename, rewrite or remove steps before running it.

Is the same prompt always planned the same way?

Clear requests are matched by rules, so the same wording plans the same way. Ambiguous clauses are classified by a model at temperature 0, which is highly consistent but not guaranteed identical.

Is this an AI agent?

No. Each step is a bounded operation (usually one model call; a research step uses a search model and a writing model), run in dependency order. That keeps runs predictable and priced up front.

See the plan before it runs.

Run the Vendor Decision Briefing in the live demo, no sign-up needed.

Start Free Try Live Demo