What you will be able to do
- Recognise when a deliverable is too large or multifaceted for a single request and split it into ordered steps
- Sequence a multi-step workflow so each step consumes the checked output of the one before
- Place validation checkpoints between steps and choose what each one checks against
- Match the decomposition pattern to the task type: fixed chain, independent sections, routing, or a feedback loop
- Carry a staged workflow across follow-ups, fresh chats and a Project without losing approved decisions
Key concept
Task decomposition (prompt chaining) — Splitting a large or multifaceted deliverable into an ordered set of smaller subtasks, each handled by its own focused prompt, with the checked output of one step becoming the input of the next.
1.When one request is too much
A single request for a big deliverable usually fails quietly. Ask for a 15-page onboarding guide in one go, built from product notes, a support-ticket export and a style guide, and you tend to get something fluent and generic that skips half the ticket themes. Nothing is broken in the prompt's wording. The request is asking Claude to plan, select, structure and write all at once, and each of those jobs gets a thinner share of attention. The fix is to break the task down: each prompt covers a single subtask, so each one has a smaller target and its output is more consistent.
Decomposition starts with you, not with Claude. You split the problem first, then decide what Claude handles at each step and where your own judgement goes in. In a long document, the step that most needs that judgement is the structure. A proven pattern is to have Claude write an outline, check that the outline meets your criteria, and only then write the document from it. For the onboarding guide, that means Claude drafts an outline from the notes and tickets, you revise it until every ticket theme has a home, and then you ask for one chapter at a time with the approved outline attached as the reference.
The tempting alternatives all keep the single request. A grander role plus a word-count floor just makes the prompt longer, and longer prompts are not automatically better. Extended thinking fixes a different problem: it supports step-by-step reasoning inside one response, while a complex multi-stage task calls for prompt chaining. Putting the ticket export into a Project's knowledge base makes the material available, but the request is still one 15-page draft with no point where you approve the structure.
| If you need… | Use… |
|---|---|
| Step-by-step reasoning | Extended thinking or chain of thought |
| Complex multi-stage task | Prompt chaining |
| Specific output format | Examples, prefilling, or explicit format instructions |
| Fewer made-up facts | Permission to say "I don't know" |
Checkpoint 1 of 8· Match them up
Match each situation to the technique that addresses it.
Tap a term, then the definition that fits it.
Each technique fixes a different problem. Extended thinking improves reasoning inside one response, but a complex multi-stage deliverable needs prompt chaining. Chaining gives each step one job and adds a point where you review the work.
“Complex multi-stage task | Prompt chaining”Source: claude.com
Checkpoint 2 of 8· Exam question
You are a marketing manager producing a 15-page onboarding guide for a new software product. You have product notes, a support-ticket export and an approved style guide. One request for the whole guide returned a generic draft that missed several ticket themes. Which approach best applies task decomposition?
Correct answer: B — Have Claude draft an outline from the notes and tickets, revise it yourself, then request one chapter at a time with that outline attached as reference.
- A. A role, a length floor and a reminder about the tickets are still prompt tweaks to one request. The guide is produced in a single pass with no reviewed structure, so missed ticket themes stay likely and nothing is checked along the way.
- B. Correct. Splitting the work into an outline you review and then one request per chapter gives each step a narrow goal. Ticket coverage is fixed in the approved outline and can be checked chapter by chapter.
- C. A Project's knowledge base changes where the material is stored, not how the work is divided. The guide is still one large request, and retrieval gives no assurance that each ticket theme is assigned to a chapter or reviewed.
- D. Extended thinking lets Claude reason before answering, but the plan stays internal and the output is still one fifteen-page draft. There is no outline to approve and no chapter-level checkpoint where a missing theme could be caught.
2.Ordering steps and passing checked output forward
In a chain, each call works on the output of the previous one, so the order of the steps is the design. A step can only be as good as the material handed to it. That means the order follows dependencies: whatever a later step relies on has to exist, and has to be correct, before that step runs.
Most staged workflows for knowledge work follow the same order:
- Define done: the audience, purpose, required sections or acceptance criteria.
- Have Claude produce structured raw material: an extraction table, or tags applied with one fixed template.
- Check that material against the original sources and correct it.
- Have Claude aggregate or draft, using only the corrected material.
- Review the result against the definition of done, or have a colleague sign off, before anyone else sees it.
Step 1 comes first because it is the yardstick for every later check. Writing acceptance criteria before any work begins gives you and Claude a shared definition of done. Step 2 uses a fixed format so the output is consistent and easy to check. Step 3 sits before any drafting because a wrong figure that slips past it gets written into the narrative, quoted in the summary and presented to leadership.
Here is how the pattern looks for a quarterly business review. Define the review's audience and sections. Ask Claude to pull the key metrics from the source reports into a table. Compare that table line by line with the reports and fix any mismatch. Ask for the narrative using only the verified figures. Then review the draft against the brief before it goes to leadership. A research memo built from 40 interview transcripts has the same shape. Claude tags each transcript using one template. You spot-check a sample of tags against the transcripts and correct the errors. Claude groups the corrected tags into themes with counts and drafts recommendations from the confirmed themes. A colleague then checks the recommendations against the transcripts before the memo is shared.
The grouping step counts tags. If tagging errors go uncorrected, they inflate or hide themes, and every recommendation built on those counts carries the error. Correcting the tags first means everything downstream works from confirmed input.
Checkpoint 3 of 8· Check yourself
A procurement lead is building a risk report from 12 vendor contracts. Claude first pulls the renewal dates and penalty clauses into a table, then writes the risk narrative from that table. Where does the most important checkpoint belong?
Each step builds on the output of the step before it. A wrong date or clause that gets past the table is written into the narrative, so the check belongs at that handoff and compares the table with the original contracts. A review at the end, whether by you or by Claude, catches errors only after they have spread through the draft.
“Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.”Source: www.anthropic.com
Checkpoint 4 of 8· Exam question
A chief analyst is using Claude to prepare a quarterly business review from several source reports. The review will be shown to leadership once it is finished. Which sequence of steps should the analyst follow?
Correct answer: C — Define audience, purpose and required sections → Extract key metrics into a table → Compare table with source reports → Draft the narrative sections → Review draft against the brief
- A. The table is compared with the source reports only after the narrative is drafted, so any extraction error is already written into the text and the draft has to be reworked instead of being built on checked figures.
- B. Metrics are extracted before the audience, purpose and sections are defined, so nothing says which metrics the review needs and the table may miss some or include ones that do not belong.
- C. The brief decides which metrics to extract, the table is checked against the sources before any text is built on it, and the finished draft is tested against the brief last.
- D. The table is compared with the source reports after the draft has already been reviewed, so a mismatch found at that point means the narrative and its review have to be redone, and leadership could receive a draft built on unchecked figures.
Checkpoint 5 of 8· Exam question
A researcher has 40 interview transcripts and needs a recommendations memo. Each step must use the confirmed output of the previous step, and a colleague signs off before the memo is shared. Which sequence should a careful researcher follow?
Correct answer: D — Tag each transcript with a fixed template → Spot-check sample tags against transcripts → Group corrected tags into counted themes → Draft recommendations from confirmed themes → Colleague reviews recommendations before sharing
- A. Grouping tags into themes happens before the spot-check, so the themes and counts are built from tags that still contain errors. Correcting the tags afterward leaves the grouped themes out of date, and the recommendations rest on unconfirmed output.
- B. The spot-check is moved to after the recommendations are drafted. Tagging errors are therefore found only after the themes, counts and recommendations have been built on them, which breaks the rule that each step uses the confirmed output of the previous step.
- C. The spot-check is pushed to the very end, after the colleague has already reviewed the recommendations. The colleague signs off on work built on unverified tags, and any correction then invalidates both the review and the memo.
- D. Tagging comes first, the spot-check corrects errors before themes and counts are built, and recommendations are drafted only from confirmed themes. The colleague's review is the final gate before sharing.
3.Validation checkpoints between steps
A validation checkpoint is a deliberate pause where an intermediate output is compared with something outside the chain before the next step uses it. It is one of the main reasons to chain at all. A single prompt hides its intermediate work, while a chain exposes it at every handoff.
What a checkpoint compares against decides whether it is worth anything. Extracted data should be compared with the source documents, and a draft should be compared with the brief or the acceptance criteria written in step 1. Each criterion needs to be concrete enough that a stranger could say whether the output passes or fails. "A patient with a basic smartphone can find the wait time in under 30 seconds" is a criterion you can check. "Easy to use" is not. Where the volume is high, as with 40 transcripts, a sample spot-check followed by corrections is a legitimate checkpoint. Re-reading everything is not required for the check to count.
Claude can run some checkpoints itself. One chain pattern is draft, review, revise: first a summary of a paper, then a prompt to review that summary for accuracy, clarity and completeness and give graded feedback, then a prompt to improve the summary using that feedback. A simpler version is to end a prompt by asking Claude to verify its answer against stated test criteria. This catches errors reliably, especially in coding and maths. Claude Opus 5 is the exception, because it already verifies its own work well without being told to. A verification line copied from a prompt that was tuned for an earlier model, one that needed the reminder, makes Opus 5 check again what it has already checked. That over-verification adds tokens and latency without adding accuracy, so on Opus 5 remove these instructions rather than rewriting them. The rule follows the model the prompt will actually run on: guidance that names a model was measured on that model, so keep the self-check line for earlier models and re-test it against your own examples before carrying it to another. Self-review supplements the human checkpoints but does not replace them. The step that confirms output against the original sources, and the sign-off before sharing, remain yours. Deciding how deep that human review should go is covered in review depth and escalation.
Checkpoint 6 of 8· Check yourself
Four checkpoints are planned for different chains. Which one cannot reliably tell you whether the output passes or fails?
A checkpoint is only useful if what it checks against is concrete. "Clear and easy to use" is a vague criterion: two people could disagree about whether the guide passes. A sample spot-check with corrections is a valid checkpoint for high-volume work. Comparing extracted data with the source reports is the standard check at a handoff. On earlier models, asking Claude to verify its answer against stated tests is a valid self-check that adds to the human checks.
“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”Source: academy.claude.com
4.Choosing a decomposition pattern by task type
Not every large task breaks down into a straight line. The shape of the work decides which pattern fits. A fixed sequence suits tasks that split cleanly into known subtasks, such as writing marketing copy and then translating it. Its trade-off is extra latency in exchange for accuracy, because each call becomes an easier task. Other shapes call for other patterns.
| Pattern | How the work is split | Fits when |
|---|---|---|
| Prompt chaining | Fixed sequence; each step processes the previous output | Subtasks are known in advance and depend on each other (outline, then document) |
| Parallelization: sectioning | Independent subtasks run side by side, then combined | Parts do not depend on each other, or each consideration deserves focused attention |
| Parallelization: voting | Same task run several times for diverse outputs | You want higher confidence, e.g. several reviews of code for vulnerabilities |
| Routing | Input is classified, then sent to a specialised follow-up | Distinct input categories are better handled separately (refunds vs technical support) |
| Orchestrator-workers | A central model breaks the task down at run time and delegates | You cannot predict the subtasks in advance, e.g. multi-file code changes |
| Evaluator-optimizer | One call generates, another critiques, in a loop | Clear evaluation criteria exist and refinement measurably helps, e.g. literary translation |
The deciding question is whether the steps depend on each other. A stakeholder-impact analysis that covers customers, employees and investors separately can be sectioned. Each audience gets its own focused request, and you merge the results. A business review cannot be sectioned that way, because the narrative depends on the verified figures, so it has to be chained. When the task has several distinct considerations, models generally do better when each consideration gets its own call.
Task type also changes how each step is prompted. Steps with an objective answer, such as code, calculations or extracted figures, suit an explicit verify-against-criteria instruction. Writing steps such as narratives and recommendations need a reference to hold on to: the approved outline, the confirmed themes, the brief. Very long technical work, like a large coding task, goes best when it is broken into increments. Claude's latest models stay oriented across long tasks by making steady progress on a few things at a time rather than attempting everything at once.
Checkpoint 7 of 8· Match them up
A consultant has four multi-part jobs this week. Which decomposition pattern fits the shape of each one?
Tap a term, then the definition that fits it.
What decides the pattern is how the parts depend on each other. The three audience write-ups don't depend on one another, so they can run side by side and be merged. The commentary depends on the verified figures, so those steps have to be chained. Inputs that fall into distinct categories call for routing. When you can't predict the subtasks in advance, a central model should split the work at run time.
“Sectioning: Breaking a task into independent subtasks run in parallel.”Source: www.anthropic.com
5.Carrying stages across follow-ups, chats and Projects
A chain does not need automation. In a chat, each stage is simply the next follow-up prompt: "Here is the corrected table; draft the narrative using only these figures." That kind of follow-up moves the plan forward. It is different from a follow-up that repairs a weak answer, which is part of prompt diagnosis and iteration. The approach holds until the conversation gets long. Claude does not learn from your earlier corrections. It only responds to what is currently in its context, and in a very long chat, earlier approved decisions can drift out of reliable reach. That is when a launch-plan chat starts contradicting pricing that was settled thirty messages earlier.
The answer is to re-supply critical context instead of hoping the model remembers it. Before you start the next stage, write a short summary of what is approved and carry it into a fresh chat. Working through the stages in chunks, with the important material at the front, is a context-as-leverage strategy. How context windows behave in detail is covered in context and memory management.
Projects give a staged deliverable a permanent home. A Project is a self-contained workspace with its own chats and knowledge base, plus instructions that shape every chat inside it. A practical setup is one chat per stage. The source material and the style guide go in the knowledge base. When a stage finishes, its approved output (the confirmed outline, the verified metrics table) goes into the knowledge base too, so the next stage's chat starts from settled decisions and not from a long, drifting thread.
For work that runs across many sessions, a short handoff note keeps the chain intact. Freeform progress notes record what was done and what comes next. Structured formats suit status data such as which sections are approved.
// Progress notes (progress.txt)
Session 3 progress:
- Fixed authentication token validation
- Updated user model to handle edge cases
- Next: investigate user_management test failures (test #2)
- Note: Do not remove tests as this could lead to missing functionalityCheckpoint 8 of 8· Check yourself
Forty messages into a launch-plan chat, Claude's draft of the go-to-market section uses a price that you rejected and corrected early in the thread. You still have three stages left. What should you do before the next stage?
Claude does not learn from earlier corrections. It works only from what is in its current context, and in a long chat, approved decisions can drift out of reliable reach. Putting the approved decisions in a summary at the front of a fresh chat, or in the Project knowledge base, makes them part of the context again instead of relying on Claude to remember them.
“The model doesn't learn from your corrections. It only responds to what's currently in context.”Source: academy.claude.com
Sources
(8)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/increase-consistencyOfficial docs
“Break down complex tasks into smaller, consistent subtasks. Each prompt then covers a single subtask, reducing inconsistency errors across scaled workflows.”
↩︎ When one request is too much“Break down complex tasks into smaller, consistent subtasks. Each prompt then covers a single subtask, reducing inconsistency errors across scaled workflows.”
↩︎ Key concept - 2.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ When one request is too much“Write acceptance tests before code. They give you and AI a shared definition of done.”
↩︎ Ordering steps and passing checked output forward“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”
↩︎ Validation checkpoints between steps“Each must be concrete enough that a stranger could tell you whether the finished product passes or fails.”
↩︎ Exam trap 2 - 3.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“Writing an outline of a document, checking that the outline meets certain criteria, then writing the document based on the outline.”
↩︎ When one request is too much“Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one.”
↩︎ Ordering steps and passing checked output forward“The main goal is to trade off latency for higher accuracy, by making each LLM call an easier task.”
↩︎ Choosing a decomposition pattern by task type“LLMs generally perform better when each consideration is handled by a separate LLM call, allowing focused attention on each specific aspect.”
↩︎ Choosing a decomposition pattern by task type“subtasks aren't pre-defined, but determined by the orchestrator based on the specific input.”
↩︎ Choosing a decomposition pattern by task type“Writing an outline of a document, checking that the outline meets certain criteria, then writing the document based on the outline.”
↩︎ Exam trap 1“This workflow is ideal for situations where the task can be easily and cleanly decomposed into fixed subtasks.”
↩︎ Exam trap 3“Sectioning: Breaking a task into independent subtasks run in parallel.”
↩︎ Checkpoint - 4.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“Don't over-engineer: Longer, more complex prompts are NOT always better.”
↩︎ When one request is too much“Review the summary above for accuracy, clarity, and completeness. Provide graded feedback.”
↩︎ Validation checkpoints between steps“Break into multiple prompts (chaining). Each prompt should do one thing well.”
↩︎ Prediction“Complex multi-stage task | Prompt chaining”
↩︎ Checkpoint“Intermediate validation adds value”
↩︎ Prediction - 5.
“Decomposes a task into sequential subtasks, where each step builds on previous results”
↩︎ Ordering steps and passing checked output forward - 6.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“This catches errors reliably, especially for coding and math.”
↩︎ Validation checkpoints between steps“verification instructions carried over from prompts tuned for earlier models can cause over-verification, adding tokens and latency.”
↩︎ Validation checkpoints between steps“On Claude Opus 5, remove these instructions rather than rewriting them.”
↩︎ Validation checkpoints between steps“Where a technique names a specific model, treat it as measured on that model and re-check it against your own evals”
↩︎ Validation checkpoints between steps“making steady advances on a few things at a time rather than attempting everything at once.”
↩︎ Choosing a decomposition pattern by task type“Use unstructured text for progress notes: Freeform progress notes work well for tracking general progress and context.”
↩︎ Carrying stages across follow-ups, chats and Projects - 7.
“Apply context-as-leverage strategies: front-loading important material, chunking long work, re-supplying critical context”
↩︎ Carrying stages across follow-ups, chats and Projects“The model doesn't learn from your corrections. It only responds to what's currently in context.”
↩︎ Carrying stages across follow-ups, chats and Projects“The model doesn't learn from your corrections. It only responds to what's currently in context.”
↩︎ Exam trap 4 - 8.
“Projects allow you to create self-contained workspaces with their own chat histories and knowledge bases.”
↩︎ Carrying stages across follow-ups, chats and Projects“you can define project instructions for each project to further tailor Claude's responses.”
↩︎ Carrying stages across follow-ups, chats and Projects