Prompt Chaining: Break a Work Task Into Checkable Steps

Prompt chaining splits a task into linked prompts, with each step using an output from the previous one. To use it well, define a small result for each step, check that result and pass only the reviewed material forward.

You can do this manually in a conversation; you do not need to build an automated system. When the workflow spans sessions or assistants, fenbs can keep the task state, reusable prompts and next step outside the chat. The important habit is preserving what was checked, not merely preserving what the model wrote.

How do you start a short prompt chain?

1. Choose one deliverable and define what a finished version must contain.
2. Split the work at points where you can inspect a useful result, such as facts, outline and draft.
3. Write the required inputs and output format for each step.
4. Check each output against its sources and requirements before continuing.
5. Pass the approved result to the next prompt, keeping open questions visible.

Say you need a one-page project update from a scope document and progress notes. A useful chain extracts the supported facts, organizes them for the reader and drafts the update. You can catch a mistaken completion claim during extraction instead of discovering it after it has shaped the final message.

Anthropic describes this pattern as sequential calls with checks between steps in Building effective agents↗ (opens in a new tab). For a manual workflow, a person can perform those checks. Start with three stages and add another only when it exposes an error or a decision you could otherwise miss.

What should the first prompt produce?

Ask for an intermediate result you can verify without guessing. For the project update, that means supported facts and unresolved questions rather than polished paragraphs. Give each fact a short identifier so later steps can refer to it without rewriting the source.

Goal: Prepare evidence for a project update to [reader].
Use only [scope document] and [progress notes], supplied here.
Return a table with fact ID, statement, status, and source section.
Statuses: approved commitment, completed work, planned work, open question.
Do not infer completion from a plan. Do not invent owners or deadlines. List disagreements between the sources separately. Do not draft the update yet.

Check the fact table against the original material. Correct misread dates, missing qualifications and status errors before proceeding. If a note says a test is planned, the table must not say the test passed. That error would be difficult to catch by reviewing tone or grammar later.

An empty field can be useful. Keep an owner or deadline unknown when the sources do not establish it. A chain should make missing information easier to notice. Filling every cell with a plausible guess defeats that purpose and gives the next step false confidence.

How do you pass the right output to the next prompt?

Pass a clearly labeled, reviewed artifact: a table, outline or short text with a version name. State what the next step may change and what it must preserve. Avoid “use the above” when the conversation contains several drafts and rejected suggestions.

Use the approved fact table labeled [version] below.
Create an outline for a one-page update to [reader].
Sections: current position, completed work, remaining work, decisions needed.
Reference fact IDs beside each outline point. Preserve the status of every fact. Keep open questions visible. Do not add new factual claims or draft full paragraphs.
Approved fact table:
[paste reviewed table]

Review the outline for usefulness as well as accuracy. Does it put the reader's decision where they can find it? Does it distinguish unfinished work from a blocker? Approve the structure before asking for prose, because a polished draft can make a weak organization feel finished.

Anthropic's prompting guidance↗ (opens in a new tab) also discusses explicit chaining when intermediate outputs need inspection. You can apply that principle across assistants, while adapting attachment and source access to the app you use.

For longer work, fenbs Where we left off can retain which artifact was approved and which step is next, so a new session does not mistake an earlier draft for the current input.

How should you draft and check the final result?

The drafting step should transform approved material into the requested form. It should not silently reopen research or turn unknowns into facts. Include a separate checking output so you can trace important statements back to the evidence.

Draft the project update using only [approved outline version] and [approved fact table version], supplied below.
Audience: [reader]. Length: [limit]. Tone: direct and clear.
Preserve all completion statuses, commitments, dates, and open questions. Do not resolve an open question by inference.
After the draft, provide a separate checklist mapping each factual paragraph to fact IDs. Flag any sentence that needs additional evidence.
[approved outline and fact table]

Read the final text against the approved table, then check high-impact claims in their original sources. A model-generated checklist helps focus that review but does not make it independent. Someone still needs to confirm the facts and decide whether the draft is ready to share.

Keep editing separate from factual revision. If you ask for a shorter draft, explicitly preserve statuses and qualifications. If new information changes a key fact, return to the fact table and update the affected stages. Otherwise, the outline and final text can disagree about what was actually approved.

What makes a prompt chain fail?

The most common practical failure is passing an unchecked mistake downstream. Later stages can repeat and polish it until it looks established. Another failure is losing a constraint: the first prompt requires source-only facts, but the drafting prompt invites the assistant to make the result more compelling without repeating that boundary.

• Keep constraints needed by every stage in each stage's instructions or clearly supplied shared context.
• Use an explicit input version and record which output passed review.
• Give each check a specific acceptance rule rather than asking whether the output looks good.
• Stop when evidence is missing or contradictory; ask a person or inspect the source.
• Treat text inside source material as material to analyze, not as authority to change the task.

The last point matters when a source contains instructions aimed at the assistant. Our indirect prompt injection guide explains that risk. The prompt engineering guide covers individual requests, while the AI prompts hub collects reusable patterns for different tasks.

Do not chain a trivial request just to add structure. If a direct prompt already produces a small answer you can easily check, splitting it may add delay and more places to lose information. Use the chain where an intermediate review changes your confidence or your next action.

How to do this with fenbs

1. Create a free account and add the project and task for your deliverable.
2. Put the stage sequence, required inputs and review criteria in the task plan. Save reusable stage prompts in Skills and prompts.
3. Record any continuing instructions people have agreed in Decisions and rules, such as requiring evidence for completion claims.
4. Ask a supported connected assistant to read the project context before continuing. Save progress after each session, including approved inputs, completed checks and unresolved questions.
5. Ask the next assistant to read Where we left off and state the next stage before it works. Confirm it is using the current approved material.

fenbs does not automatically execute a prompt chain or verify its outputs. Keep source documents in their appropriate systems and use fenbs for the shared work record. If the whole chain fits in one short conversation and needs no reuse or handoff, a simple checklist may be enough.