Prompt Chaining Strategies
Breaking complex tasks into sequential prompts instead of one mega-prompt
When one prompt isn’t enough
Some tasks are genuinely too complex for a single prompt to handle well — not because Claude can’t process the length, but because cramming research, analysis, and writing into one instruction forces Claude to do all three at once, with no chance to course-correct between stages. Prompt chaining means splitting a complex task into a sequence of smaller prompts, where the output of one becomes the input to the next.
Why chaining beats a mega-prompt
A single giant prompt asking Claude to “research this topic, analyze the findings, and write a polished report” asks Claude to do three different cognitive jobs in one pass, with no opportunity for you to inspect or correct the middle steps. If the research is subtly wrong, that error propagates silently into the analysis and the final report — and you won’t catch it until the end, if at all.
Chaining fixes this by making each stage’s output a checkpoint. You (or Claude, in a self-critique loop) can inspect and fix problems before they compound.
Key principle: Chain when stages depend on quality control between them. Don’t chain just to make a prompt shorter — if the stages don’t need independent verification, one well-structured prompt is faster and cheaper.
A basic three-stage chain
PROMPT 1 (extract):
"Read this raw interview transcript and extract every direct quote
where the customer describes a specific pain point. Output as a
numbered list, quote only, no commentary."
→ [Claude's output: 12 numbered quotes]
PROMPT 2 (structure) — paste the 12 quotes in:
"Group these 12 customer pain point quotes into 3-5 themes.
For each theme, give it a short name and list which quote numbers
belong to it."
→ [Claude's output: themed groups]
PROMPT 3 (write) — paste the themed groups in:
"Using these themes and supporting quotes, write a 400-word
executive summary of customer pain points for our product team."Each stage has one job. If stage 2’s grouping looks wrong, you fix it before it becomes the backbone of a summary that goes to stakeholders.
Passing state between links
The cleanest way to chain is to literally copy Claude’s output from one response into your next prompt — you stay in control of what carries forward, and you can edit the intermediate output before passing it on. This is different from just continuing the same conversation, where everything accumulates and Claude has to hold the entire history in view.
For chains you’ll run repeatedly, structure each link’s output to be easy to paste into the next: numbered lists, labeled sections, or a simple format rather than free-flowing prose. A stage that outputs “Theme 1: Onboarding friction — quotes 2, 5, 9” is trivial to paste into the next prompt; three paragraphs of prose about themes is not.
Common chain patterns
| Pattern | Use case |
|---|---|
| Extract → Structure → Write | Turning raw data (transcripts, notes) into a polished deliverable |
| Draft → Critique → Revise | Improving quality on writing or analysis (see Lesson 6) |
| Outline → Expand section by section | Long documents where each section needs independent focus |
| Translate → Verify → Adjust | High-stakes accuracy tasks where a second pass checks the first |
Try it: Take a task where you currently use one long prompt and the results are inconsistent. Split it into two stages with a clear handoff, run both, and compare the result to your usual single-shot version.
Watch for: Chaining adds latency and manual steps. If a task is simple enough that a single well-structured prompt gets it right consistently, chaining just adds overhead for no benefit.