Prompt chaining: break a big job into a chain of AI prompts

Why one big prompt does each part worse

Ask for a job description, a scorecard, interview questions and an offer letter in one prompt, and you get all four, each thinner than it would be alone. One answer has to cover everything, so each part gets a share of it. The model also has to settle, silently, the decisions you would have made between the parts, such as which requirements matter most and what the budget allows. When it guesses wrong early, every later part is built on the guess, and there is no point where you could have caught it.

Anthropic’s prompting guide notes that its recent models handle most multistep reasoning on their own, and that explicit chaining is still useful when you need to inspect the answers in between. At work you nearly always do: a scorecard is a decision your team owns, not a draft the model should settle for you.

A hiring chain: each step hands one checked answer to the next

Before the interviews. IntakeA role brief agreed with the hiring managergives the brief
Before the interviews. ScorecardCriteria and questions, from the brieftakes the brief
Before the interviews. Phone screenThe same questions for every candidatetakes the scorecard
During and after them. Interview feedbackEach interviewer’s notes, against the scorecardtakes the scorecard
During and after them. DebriefAll the feedback, criterion by criteriontakes the feedback
During and after them. Offer letterFrom the terms you agreedtakes your terms

Not every step takes the one before it: the phone screen and the interview feedback both work from the scorecard.

1. Give each step one deliverable

A step is the right size when it produces one thing you would want to read before going on: a brief, a list, a table, a draft. If you would not stop to check it, merge it with the next step. If you keep fixing two separate things in one answer, split it.

Write each step as a full prompt of its own, with its reader, its format, its length and what to do when unsure. A chain of vague prompts is still vague; the guide to briefing AI like a new hire covers what each one needs.

2. Check each answer before you pass it on

The stops between steps are the point of a chain. Read each answer before it goes on and fix it there: an extra criterion in the scorecard is one edit now and a wrong question in every interview later. Look for what the next step will depend on, such as facts that are not in your material, criteria that are missing, a tone that is wrong for the reader, and anything only you can decide.

Correct the answer itself rather than only the next prompt, so that what passes on is what you approved. If a step keeps needing the same fix, move the fix into that step’s prompt.

3. Say which step takes which answer

A chain is not always a straight line. In hiring, the phone screen and the interview feedback both take the scorecard from step 2, and the debrief takes every interviewer’s feedback. The offer letter takes the terms you agreed, which no earlier step should invent. Note for each step which earlier answer it takes, and paste that answer between tags, so the model can tell it from your instructions:

Write an interview scorecard for the role in the
brief below.

<role_brief>
(paste the brief you approved in step 1)
</role_brief>

Pass on only what a step needs. The debrief needs the feedback, not the phone screen script, and every extra page is one more thing for the model to draw on in the wrong place.

4. Give shared details once

Some details turn up in nearly every step: the role title, the team, the hiring manager’s name, the salary range, the start date. Write them once as a short block and paste it at the top of each prompt, or set them in a custom GPT or Project that you run the steps in. Typing them fresh each time is how one step ends up with a different salary range from the next.

The Workflow Planner does this for you. Describe the job in a sentence, and it lists the steps in order with an expert prompt for each. It asks for every shared detail once and fills it into each prompt that needs it. When you paste a step’s answer back, the next prompt is ready with that answer inside. The details you fill in stay in your browser.

Example: hiring for a role

Here is the hiring chain from the graphic, with what each step makes, what it takes and what you check before going on.

Step	It makes	It takes	You check
1. Intake	A role brief: must-haves and nice-to-haves	Your notes from the hiring manager	That each must-have truly is one
2. Scorecard	5 to 7 criteria, each with questions	The brief from step 1	Nothing the brief does not support
3. Phone screen	One script for every candidate	The scorecard	Questions you would ask everyone
4. Interview feedback	One write-up per interviewer	The scorecard and that interviewer’s notes	Evidence behind every rating
5. Debrief	The panel’s feedback side by side	Every write-up from step 4	Disagreements shown, not averaged away
6. Offer letter	A draft for HR or legal review	The terms you agreed, typed by you	Every figure against the agreed terms

Each step has a tested prompt in the repository, from the intake meeting to the offer letter. The Workflow Planner lays them out in this order, with one more step for the email that tells candidates what to expect.

Hiring a designer: one big prompt, then step 2 of the chain

Before

Help me hire a senior product designer. Write the job description, a scorecard, phone screen questions, interview questions and an offer letter.

1/10 Needs work

Has: What to make
Missing: Who it's for
Missing: Background facts
Missing: Length
Missing: Format
Missing: What to avoid
Missing: What good looks like
Missing: An example
Missing: If unsure
Missing: Layout

After, with what was added marked

Write an interview scorecard for the senior product designer role, from the role brief below. The hiring manager and the four interviewers will use it to judge every candidate the same way. <role_brief> {{role_brief_from_step_1}} </role_brief> Include 5 to 7 criteria, taken only from the must-haves and nice-to-haves in the brief. For each criterion give: 1. What it means in this role, in one sentence. 2. Two interview questions that ask for a real example. 3. What a strong answer and a weak answer sound like. 4. A rating from 1 to 4, with what each number means. Do not add criteria the brief does not support, such as culture fit or years of experience it does not ask for. If the brief is unclear about a must-have, list it under Questions for the hiring manager instead of guessing. Format: a table with one row per criterion, then the questions as a numbered list. A good scorecard lets two interviewers who never spoke reach the same rating from the same answer.

9/10 Strong

Has: What to make
Has: Who it's for
Has: Background facts
Has: Length
Has: Format
Has: What to avoid
Has: What good looks like
Missing: An example
Has: If unsure
Has: Layout

Illustrative. The first prompt asks for five deliverables at once and leaves every decision to the model. The second is step 2 alone: it takes the brief you approved and says who uses the scorecard, and questions only the hiring manager can answer go back to them. The marked words are what it adds.

Example: interviews into a report

Say you ran twelve customer interviews and leadership wants a short report. Pasting all twelve transcripts into one prompt makes a very long input, and long inputs are where details in the middle go missing, as the guide to summarizing long documents explains. A chain keeps each step small enough to check:

One summary per interview. The problems the person raised, each with a quote and where it appears. Takes one transcript at a time. Check a few quotes against the transcript.
Themes across interviews. Each theme with how many interviews raised it and the clearest quotes. Takes the checked summaries, not the transcripts. Check that no theme rests on one loud interview.
Findings. What the themes mean for the decision in front of leadership. Takes the themes you agreed. Check that every finding traces back to a theme.
The report. One page, with the decision it supports first. Takes the findings. Check that its numbers and quotes match step 2.

The interview synthesis prompt covers the first two steps.

Try it: make this the first step of a chain

The prompt, which you can edit

2/10 Needs work

Has: What to make
Has: Who it's for
Missing: Background facts
Missing: Length
Missing: Format
Missing: What to avoid
Missing: What good looks like
Missing: An example
Missing: If unsure
Missing: Layout

Make it step 1 only: a summary of one interview, with the problems the person raised and a quote for each. Then add who reads it, the format, the length and what to do when a quote is unclear. Open in the Grader

When one prompt is enough

Not every job needs a chain. If there is nothing you would check in the middle, such as a short email or a one-paragraph summary, one good prompt is faster. Chain when the job has a decision between its parts, or when one part feeds several others.

You can also ask the model to plan the chain before you run it: describe the job and ask for the steps in order, with what each one hands to the next. Then check the plan the way you would check any other answer.

Key takeaways

One big prompt spreads one answer across every part and hides its guesses where you cannot check them.
Give each step one deliverable you would want to read before going on.
Check and correct each answer before it feeds the next step.
Note which earlier answer each step takes, and paste it between tags.
Give shared details, such as the role title and salary range, once.
Chain when the job has decisions between its parts; for a short task, one good prompt is enough.

Questions

What is prompt chaining?

Prompt chaining is splitting a task into a sequence of prompts, where the answer from one becomes part of the next. Each prompt does one part of the job, and you check its answer before it moves on. Hiring is a typical chain: the role brief feeds the scorecard, and the scorecard feeds the phone screen and the interview feedback.

Is prompt chaining better than one long prompt?

For a job with several deliverables, usually yes, because you can check and fix each part before the next one builds on it. For a short task with nothing to check in the middle, one good prompt is faster. Anthropic’s prompting guide says chaining stays useful when you need to inspect the answers in between, which is the case whenever a person has to approve a step.

How do I pass one prompt’s answer to the next in ChatGPT or Claude?

Copy the answer you approved and paste it into the next prompt under a clear label, or between tags named for what it is, with a line saying what it is for. A new chat for each step means the model sees only what you pass it; one long chat is quicker, but drafts you rejected stay in view. The Workflow Planner puts each answer into the next prompt for you.

How many steps should a prompt chain have?

One for each answer you would want to check before going on. Hiring for a role takes about six; turning interviews into a report takes four. If you would never read a step’s answer on its own, merge it with the next; if you keep fixing two separate things in one answer, split it.

Can an AI agent run the whole chain for me?

Agents can work through several steps on their own. What you lose is the stop between steps where you check the answer. For work where a person must approve the middle, such as a scorecard or a set of findings, run the steps yourself or tell the agent to stop and show you each answer before it goes on.

Prompts to use with this

Prepare a recruiter intake meeting with a hiring managerA new role has opened and you need the right questions for the hiring manager and a role brief you both agree on.
Build a structured interview plan and scorecard for a roleYou are about to interview for a role and want every interviewer asking job-related questions and scoring answers against the same bar.
Run a hiring debrief from the interview panel's feedbackThe panel has interviewed a candidate and you need their feedback organized by criterion so the hiring manager can decide fairly.
Synthesize user interviews into findingsYou have run several customer interviews and need findings backed by counts and quotes, not the loudest anecdote.

Browse all 669 prompts