Prompt Chaining: 5 Recipes That Ship (2026)

Quick answer

Prompt chaining is splitting one job across two or more prompts, where each prompt's output feeds the next input. In 2026 it buys you accuracy and control on multi-step tasks, at the cost of latency, tokens, and more places to fail. Below are five paste-ready chains, each with the failure mode nobody warns you about, plus a section on when a single prompt wins.

Both Anthropic's chain-prompts guidance and OpenAI's prompting guide say the same thing: when a task has distinct steps, give each step its own prompt. The community reference promptingguide.ai frames it as trading one opaque call for a pipeline you can inspect. The recipes below are that pipeline. Terse, paste-ready.

Recipe 1: Extract, then reason

Never make one prompt extract and decide at once.

Prompt A pulls structured facts. Prompt B reasons over only those facts.

[Prompt A] Extract every date, amount, and party from the contract below as JSON. No commentary.
[Prompt B] Given this JSON, list every payment due in the next 30 days.

Why it works: a model that reasons while reading invents facts. Split it. Prompt B never sees the raw text, so it cannot smuggle in a clause that was not there.

Fails when: Prompt A drops a field B needs. Garbage in, garbage out, and B sounds confident about it. Test A alone first.

Recipe 2: Draft, critique, rewrite

Two passes beat one long instruction.

[Prompt A] Write the release note. 120 words.
[Prompt B] You are a critic. List 3 concrete problems with this note. No praise.
[Prompt C] Rewrite the note fixing exactly these 3 problems.

Why it works: a fresh context critiques harder than the writer. The model will not defend a draft it does not remember writing. This is the meta-prompting move aimed at your own output; there are more of these in the meta prompting recipes.

Fails when: the critic invents nitpicks to look useful. Cap it at 3 and force the word "concrete."

Recipe 3: Route, then answer

Classify first. Dispatch to a specialized prompt second.

[Prompt A] Classify this ticket: BILLING, BUG, or HOWTO. One word only.
[then] send it to the prompt written for that class.

Why it works: one giant do-everything prompt is worse at all three jobs than three small prompts are at one each. Routing keeps every instruction short and testable. It also pairs cleanly with tool calls, covered in the tool-use recipes.

Fails when: the classifier returns "BILLING/BUG" or a whole sentence. Constrain output to the exact enum and reject anything else.

Recipe 4: Map, then reduce

For long input, summarize the parts, then merge.

[Prompt A, per chunk] Summarize this section in 3 bullets.
[Prompt B] Merge these bullet sets into one 5-bullet summary. Drop duplicates.

Why it works: it beats stuffing 200 pages into one call. Each chunk gets full attention; the merge sees only signal.

Fails when: cross-chunk context matters, like a name defined in chunk 1 and used in chunk 9. Map-reduce loses it. Add a shared glossary pass before the map step.

Recipe 5: Generate, then gate

Add a validator hop that returns pass or fail with a reason.

[Prompt A] Generate the SQL for this request.
[Prompt B] Return PASS or FAIL and one reason. FAIL if it writes, drops, or lacks a LIMIT.
[on FAIL] re-run A with the reason appended.

Why it works: cheaper and more honest than trusting the first output. The gate is a second opinion you can log and count.

Fails when: A and B share the same blind spot. Use a different instruction, and sometimes a cheaper model, for the gate so it fails differently than the generator.

When a single prompt wins

Chains are not free. Each hop adds latency, tokens, and a new failure surface. Every hop also drops context the last one had.

Skip the chain when:

• The task is one step. Classification, extraction, a short rewrite. Sections inside one prompt do the job.
• Latency matters. A 4-hop chain is 4 round trips. Users feel it.
• You are in a loop or a hot path. Chain cost multiplies on every call.

Chain when the task genuinely has stages that need to be verified or specialized between them. Not before.

Cost to test: $0.03 to run all five chains once on a small 2026 model, most of it the map-reduce pass.