Chaining AI outputs into a pipeline

The output of step one becomes the input to step two

After this, you'll be able to take a complex task and split it into a sequence of AI calls where each step's output becomes the next step's input, producing better results than a single prompt could.

Before you start

Complete Context Scaling first; this lesson builds on managing context and state across multiple AI interactions.

The idea

When you ask an AI to research, analyze, and write a report in a single call, it splits its attention across three tasks and does all three worse than it would if focused on one. The fix is compounding engineering: splitting the work into sequential calls where each output feeds the next. The loop is simple: extract, structure, compose.

Here is the before and after: a single prompt ('Turn my meeting notes into a follow-up email') produces a generic summary that misses specific decisions and action items. A three-step pipeline produces a more specific result. Step 1 extracts raw action items from the notes as a bulleted list. Step 2 takes that list and groups items by owner with deadlines. Step 3 takes the organized list and drafts a follow-up email using only those specific items. Each step is focused, each output is better, and the final email contains the actual decisions from your meeting, not a generic recap.

Now try it: take a task you currently handle in one long prompt and break it into two steps where the output of step 1 is pasted as input to step 2. The discipline is resisting the urge to combine them back into one step after it works.

Try it (12 min)

Watch out for

Passing the raw step 1 output to step 2 without reviewing it. Errors in step 1 compound into step 3 and the final output looks plausible but is wrong.
Collapsing a two-step pipeline back into one when the first attempt produces mediocre output. Mediocrity usually means a step is still doing too much at once, not that the pipeline is too complex.
Building a chain of more than four steps before testing; test steps 1 and 2 together before adding step 3
Forgetting to add [bracketed placeholders] in your templates; without them the pipeline is locked to one example and cannot be reused

Paste this into Claude

I want to build a 3-step AI pipeline for [describe the end goal, e.g., turning a rough brain dump into a formatted client proposal]. Step 1 should [describe what the first call should extract or identify from my raw input; be specific about the output format]. Step 2 should [describe how to structure or transform step 1's output; name the grouping or organization method]. Step 3 should [describe the final output that uses the structured content from step 2]. For each step, write a reusable prompt template with [bracketed placeholders] for the content that changes each time. Then run step 1 on this example input and show me the output I would paste into step 2: [paste a real example of your raw input material].

What good looks like

Each step has its own prompt template with clear [bracketed placeholders] for the changing content
Step 2's prompt explicitly includes a slot for 'the output from step 1' as its input
Running step 1 on your example produces an output you would actually paste into step 2 without editing
Run the same raw input through a single prompt and compare. The pipeline output contains specific names, decisions, and structure that the single-prompt version missed.

When this breaks

Breaks when step 1's output format changes unexpectedly because step 2 assumed a specific structure and now generates incorrect content to fill the gap
Breaks when intermediate steps are not tested because errors accumulate silently across the chain and the final output is wrong in ways that are hard to trace back

You can now

Run the pipeline on a second piece of input and confirm each step produces useful output without you manually editing the content between steps.

Key takeaways

One long prompt produces mediocre output across three tasks. Three focused prompts, each passing output to the next, produce better output at every step.

Compounding engineering means the output of one AI call becomes the structured input to the next
Extract, structure, compose is the simplest three-step pipeline for any raw-notes-to-document task
Test each step independently before connecting them; errors in step 1 compound through step 3
The discipline is keeping steps separate even after the pipeline is working

Was this helpful?

← Back to lessons

I want to build a 3-step AI pipeline for [describe the end goal, e.g., turning a rough brain dump into a formatted client proposal]. Step 1 should [describe what the first call should extract or identify from my raw input; be specific about the output format]. Step 2 should [describe how to structure or transform step 1's output; name the grouping or organization method]. Step 3 should [describe the final output that uses the structured content from step 2]. For each step, write a reusable prompt template with [bracketed placeholders] for the content that changes each time. Then run step 1 on this example input and show me the output I would paste into step 2: [paste a real example of your raw input material].