TL;DR
Prompt chaining splits a huge task into a sequence of small, dependent prompts — each output becoming the next input — instead of demanding the finished product in one impossible breath. Quality compounds down the chain: errors get caught at the station where they're cheap, and each step starts from verified material instead of a guess.
The Analogy
There are two ways to get a car.
The first is the Magic Box: a warehouse with a single enormous machine. Raw ore and rubber go in one end; a finished car rolls out the other. It's a seductive idea — one step, zero effort. It's also why nobody has ever built one. A machine that must simultaneously smelt steel, blow glass, sew seats, and calibrate brakes can be optimized for none of them, and when it produces a defective door, the entire car is scrap — because the flaw was born in step one and discovered in step fifty.
The second is the assembly line: dozens of stations, each doing exactly one thing supremely well. Steel is stamped here, and inspected before it moves. Doors are welded there, and checked before painting. When a defect appears, it's caught at its station — after a few operations, not after the whole car. Quality control isn't a final gamble; it's built between every pair of steps. That's not the line's weakness, being "slower" — it's the entire reason cars became reliable and affordable at all.
Asking AI for a complex deliverable in a single mega-prompt — "research my market, write the report, design the slides, and draft the announcement email" — is an order for the Magic Box. The model must juggle four expertises at once, and it optimizes for none of them. Worse: an error made during the invisible "research" phase surfaces only in the final email, fully dressed as finished work, with every subsequent step built on top of the mistake.
Prompt chaining is you building the assembly line instead. Small stations. One operation each. Inspection between steps. The conveyor belt between stations already exists — it's the conversation itself.
How It Works
Every productive chain follows the same three-station backbone — Gather → Transform → Polish — extended with as many intermediate stations as the job demands.
| Station | Job | Example prompt |
|---|---|---|
| Gather | Produce raw material: facts, options, research | "List the 8 most common objections to electric vehicles, one line each" |
| Transform | Convert material into structure: outline, draft, skeleton | "Group those objections into 3 themes and outline a section per theme" |
| Polish | Finish one station's output at a time: style, tighten, audit | "Write section 1 only, 150 words, confident tone" |
Three mechanics make the chain work:
1. Each station inherits everything. The model re-reads the whole conversation before every reply — that's the context window, the desk from lesson two. Step three doesn't need you to re-paste the outline from step two; it's already on the desk, along with the objections from step one. The chain's memory is the conveyor belt. (Which is also why chains live in one chat per project: the belt only connects stations on the same line.)
2. Each station is small enough to verify. An eight-line list of objections can be checked in ten seconds — is #6 actually a real objection? Catch it there, and the outline never inherits it. This is the quality magic: errors compound in a mega-prompt but get intercepted in a chain. You've already met the two-station version of this idea — the Auditor Prompt is exactly a chain: station one drafts, station two attacks, station three fixes. Chaining generalizes that rhythm to any scale.
3. Each station gets its full attention. A model asked to do one thing gives that thing everything it has. The same model asked to do four things at once is negotiating a budget between them internally — and the polish station is always the first casualty of the mega-prompt. On a line, every station is the model's only job.
The working rule:
One station, one operation, one inspection. If a step's output couldn't be checked in under a minute, the station is too big — split it.
Before & After (The Prompts)
Example 1 — the market report
Research the home-fitness market and write a complete 1,500-word report with an executive summary, competitor analysis, three growth opportunities with supporting arguments, and a conclusion. Make it professional and persuasive.
The Magic Box. What returns is shaped like a report, but the reasoning is invisible — you can't tell which "growth opportunity" rests on a fact the model invented under all that simultaneous load. When opportunity #2 turns out to be hollow, you don't know whether the research failed or the writing did. Rebuild from scratch.
Step 1: List the 5 major home-fitness competitors with their price point and one differentiator each, as a table. If unsure about any figure, mark it [VERIFY].
Check the table. Two minutes. Fix the one [VERIFY] that matters, and the raw material is now known good — every later station stands on it.
Step 2: Based on that table, outline a 4-section report: one section per growth opportunity you can defend from this data, plus an executive summary. Section titles only, one line of argument each.
Check the outline. Is opportunity #3 actually defensible from the table, or is it vibes? Kill it now — before prose exists to make it look smarter than it is.
Step 3: Write the executive summary only — 120 words, confident, no claims that go beyond the table above.
One station's worth of full attention, built on verified structure. Then continue down the line, one section per prompt, in any order you like. The final report isn't luck; it's the last station of a line where every earlier station passed inspection.
Example 2 — the trip plan
Plan a perfect 7-day trip to Portugal for a family with two kids, including cities, hotels, restaurants, and daily activities.
One mega-prompt, five expertises (logistics, budget, child-friendliness, pacing, food), zero inspectable intermediates.
Step 1: We're a family with kids aged 6 and 9, first time in Portugal, 7 days, mid-budget. Ask me up to 5 questions — one at a time — about anything that changes the plan.
The interview as station one: the gather station for facts only you hold. Then:
Step 2: Given my answers, propose 3 itinerary skeletons — different city sequences and pacing. Table format, one row per day, no details yet.
Three cheap skeletons, checked in a minute: the Algarve-first route dies because of the June heat you mentioned in step one — caught at the skeleton, not on day four of the vacation. Only then does step three flesh out the winner, day by day, one station per day.
Example 3 — the job application
Write a cover letter for this product-manager job using my résumé.
The classic one-breath request — and the reason most AI cover letters sound like everyone else's AI cover letter.
Step 1: From the job posting below, extract the 5 requirements they care about most, ranked. Quote the posting for each. Step 2: From my résumé, match my strongest evidence to each requirement — one row per requirement, quote my résumé. Step 3: Write the cover letter using only the five matched pairs above. 250 words, no clichés, open with my strongest match.
Station one checks: are these really the top requirements, or the obvious ones? Station two checks: do I actually have the evidence, or am I about to overclaim? Only then does station three write — constrained, grounded, and unrecognizable from the default draft.
Common Pitfalls
- Chaining too fast — never inspecting the intermediates. A chain whose stations are never checked is just a slow mega-prompt with extra steps. The value isn't the sequence; it's the inspection points between stations. Pause at every output: ten seconds now, or a full rebuild later.
- Stations that are secretly mega-prompts. "Step 2: write sections 1 through 4" is the Magic Box hiding inside your line. One station, one operation — if you couldn't verify its output in a minute, split the station.
- Letting the chain sprawl. Twelve stations for a two-paragraph email is ceremony, not craft. Chains are for genuinely multi-stage work; small jobs want a single well-aimed one-shot prompt. Build a line when the product has parts; don't build one when it has a sentence.
- Spreading the chain across multiple chats. The conveyor belt is the conversation's memory — open a fresh chat for "step 3" and the desk is empty: no outline, no interview answers, nothing. One project, one chat, top to bottom.
- Restarting the whole chain after one bad station. A flawed step-two outline doesn't send you back to the raw ore. Fix that station's output ("regenerate the outline, drop section 3") and continue — the line resumes from the last good station. That's precisely the repair speed the Magic Box never offers.
FAQ
What is prompt chaining?
Prompt chaining is decomposing a large AI task into a sequence of small, dependent prompts — gather, transform, polish — where each step's output becomes the next step's input within the same conversation. Instead of one mega-prompt producing an unauditable blob, you get stations you can inspect and correct individually, so errors are caught where they're cheap and quality compounds down the line.
Why does chaining give better results than one big prompt?
Three compounding reasons: each step gets the model's full attention instead of a share of it; each intermediate output is small enough to verify, so mistakes are intercepted at their station instead of discovered in the final product; and every downstream step is built on inspected material. The Auditor Prompt — draft, then attack — is the two-station proof of the same principle.
How many steps should a prompt chain have?
As many stations as the work has inspectable parts, and not one more. A report might need five; a cover letter needs three; a sentence needs one. The test isn't a count — it's whether each station performs a single operation whose output you can check in under a minute. When in doubt, split the station that feels too big and delete the station nobody inspects.
Next Lesson
Chaining controls sequence — the order in which the AI works. But there's a second dimension the chain can't touch: style. How do you make the model's output look exactly like the output you have in your head? Not by describing it — by showing it. One worked example in the prompt locks a pattern tighter than any instruction ever written.
Continue to the next lesson: Few-Shot Prompting
Practice drill before you go: take the next big deliverable on your list and write only station one — the gather step — before touching the AI. Define what its output looks like, ship it, and inspect what comes back. Feel the difference between hoping the final product is right and knowing its foundation is.