What you will be able to do
- Put the five stages of a research, analysis or planning deliverable in order and explain why each one sits where it does
- Write a requirements brief before any research or drafting starts, and recognise shortcuts that skip it
- Run research that cites its sources, and check that the key claims are actually supported
- Use code execution on uploaded data and trace an implausible figure back to its cause in the right order
- Turn verified findings into an actionable plan that a person reviews before it is shared
Key concept
Staged deliverable workflow — A finished deliverable is built as a chain of stages: brief, cited research, code-based analysis, synthesis and human review. Each stage's output becomes the next stage's input, so every intermediate result can be checked before anything is built on it.
1.Why a deliverable is built in stages
A market-entry plan, a vendor review or a budget proposal is never a single prompt. It is a sequence of stages, and each stage's output becomes the input to the next. Running them as separate steps lets you inspect each intermediate result before it feeds anything else. In a single pass, a mistake made early carries through to the end without anyone seeing it. The most common chain is draft, then review against criteria, then refine, with each step run separately so you can log, evaluate or change direction at any point.
For a research, analysis or planning deliverable, the stages run in this order, and each placement has a reason:
- Requirements brief. It sets the audience, the decision the deliverable supports, the scope limits and the success criteria. Every later stage is judged against it, so it comes first.
- Research with cited sources. It runs against the brief. A person opens the sources behind the key claims. The research comes before the data work because it shows which questions the numbers need to answer.
- Data analysis through code execution. Claude computes the figures the plan depends on from the team's own uploaded data. It is designed around the questions the research surfaced.
- Synthesis. Claude turns the verified findings and figures into a plan with actions, owners, dates and risks.
- Human review. A team member checks the plan against the brief, and checks its figures, before it leaves the team.
| Claude runs code for | Claude answers directly for |
|---|---|
| Non-trivial math (large numbers, many steps, precision-sensitive results) | Simple arithmetic and well-known math facts |
| Data analysis, file parsing, or visualization | Factual, conversational, or creative requests |
| Explicit requests to run, compute, or execute | Simple unit conversions or translations |
Checkpoint 1 of 8· Check yourself
A manager asks Claude for a complete vendor-review plan in one prompt: research, figures and recommendations all at once. What is the main reason to run the work as separate stages instead?
Chaining stages as separate calls exposes each intermediate output for checking. In a single pass, an early error flows through to the final deliverable without anyone seeing it.
“Explicit prompt chaining (breaking a task into sequential API calls) is still useful when you need to inspect intermediate outputs”Source: platform.claude.com
Sources1
2.Requirements analysis: the brief comes before the draft
Claude starts every deliverable like a brilliant new hire who knows nothing about your organisation. It has no way of knowing who in leadership will read the plan, which decision the plan feeds, or what budget and deadline apply. Only the person who asked holds that information. The more precisely it is spelled out, the better the result. Explaining why you want something also helps: it gives Claude more information about your goals and leads to more targeted output.
A requirements brief records four things:
- Audience: who will read the deliverable and act on it.
- The decision it supports: for example, whether to consolidate vendors in one region, not "everything about vendors".
- Scope limits: spending boundaries, deadlines, and the regions or products that are in or out.
- Success criteria: what a good result looks like.
Success criteria need to be specific, so they say exactly what you want to achieve, and measurable, using numbers or clearly defined qualitative scales. A brief that asks for "a good plan" gives the final reviewer nothing to check against.
Suppose the request is only two lines: "Build a vendor consolidation plan for leadership." The right move is to ask about audience, decision, limits and success criteria first, then draft. Each common shortcut fails for a specific reason:
- Drafting on stated assumptions and letting the manager mark it up. This guarantees a rewrite once the assumptions turn out wrong.
- Researching how other companies did it and copying the common structure. This gives a format without knowing what the format is for.
- Asking for contracts and spend records and inferring priorities from them. Data shows what the company spends, not which choice leadership faces or what limits it has set.
The same applies to a vague research request such as "tell me everything about remote work trends". The fix is to name the actual decision (keep a hybrid schedule for 300 office staff in Germany or not) and let that decision set the scope.
Checkpoint 2 of 8· Check yourself
Two lines arrive from a department head: "Build a vendor consolidation plan for leadership. Need it soon." Nothing else is attached. What should happen before Claude writes any of the plan?
Claude starts without the requester's context. The audience, the decision, the limits and the success criteria are known only to the person who asked, so gather them before drafting. A draft built on assumptions leads to a rewrite. Spend data shows what the company spends, not what leadership has to decide. Copying a peer's structure gives you a format without its purpose.
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”Source: platform.claude.com
Checkpoint 3 of 8· Exam question
A regional operations manager sends Claude a two-line request: "Build a vendor consolidation plan for leadership." The manager has not said who in leadership will read it, which decision it supports, or what limits apply. What should happen before Claude drafts anything?
Correct answer: B — Claude asks the manager about the intended audience, the choice the plan feeds, spending and deadline boundaries, and success criteria, then drafts.
- A. Listing assumptions in a draft does not replace getting the requirements. The manager has not said who will read the plan, what decision it supports or what limits apply, so a full draft is built on guesses about all three. Reworking it afterward costs more than settling them before drafting.
- B. Requirements analysis comes before drafting. Audience, the decision being supported, constraints and success criteria shape every later stage, and only the manager can supply them. Getting them first avoids a polished draft aimed at the wrong target.
- C. Outside examples cannot reveal who will read this plan or what decision it supports. The most common structure across other companies may not fit these readers or limits, and research is only useful once the requirements say what to look for. Reviewing the outline afterward does not fix a starting point chosen without them.
- D. Contracts and spend records show what the company owes and pays its vendors, not who in leadership will read the plan, what they must decide or what limits they set. Those details have to come from the manager rather than be inferred from records, so structuring the plan from them is guesswork.
3.Research with cited sources
Research is where Claude reaches information outside its training. With web search, Claude can answer from current content beyond its knowledge cutoff, and its final response cites the sources it drew on. Claude searches when a request depends on information that is current, changing, or outside its training data: recent announcements, current prices, or details about organisations that may have changed. It does not search for stable facts or for analysis of material already in the conversation. Because the brief defines the decision, the research query can be targeted rather than open-ended.
Research rarely follows a fixed path. Researchers change their approach as they find things and follow leads that come up along the way. For a complex question, a structured prompt helps. Ask Claude to develop competing hypotheses as it gathers data, track its confidence, critique its own approach, and keep research notes. This helps it work through large amounts of material methodically.
Citations make the research auditable, but they do not prove it is right. Asking Claude to cite quotes and sources for each claim lets a person open the source behind every key claim and check that it actually says what the claim says. How to judge that support is covered in hallucination detection and verification.
Web search is not the only way to get citations. Claude can also cite documents you supply, such as PDFs or plain text files, when citations are enabled on them through the API. A common use is asking it to pull specific fields from those documents, for example a price or a date; each field's answer is an extracted value. Two practical pitfalls apply to that case:
- Missing citations. When citations are produced through the API, check that each extracted value has at least one citation, because the API does not guarantee one.
- JSON-only prompts. A prompt that asks for nothing but JSON can come back with few or no citations at all.
Checkpoint 4 of 8· Check yourself
Web-search research for the vendor consolidation plan comes back with a citation on each key claim. What should happen to those claims before they feed the analysis stage?
Citations make research auditable, but they do not show that it is right. Someone still has to open the source behind each key claim and check that it supports the claim. A JSON-only prompt makes things worse because it can return few or no citations.
“Make Claude’s response auditable by having it cite quotes and sources for each of its claims.”Source: platform.claude.com
4.Data analysis through code execution
Figures the plan depends on should be computed, not estimated. With the code execution tool, Claude writes and runs Python and Bash in a sandboxed container. It can analyse data, run calculations and produce charts. The container has no internet access, so the team's own CSV or Excel files have to be uploaded through the Files API and referenced in the request. For a borderline request where Claude might answer from memory, ask for computation explicitly, for example "run code to verify this".
# Upload a file
file_object = client.files.upload(file=Path("data.csv"))
# Use the file_id with code execution
response = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "Analyze this CSV data"},
{"type": "container_upload", "file_id": file_object.id},
],
}
],
tools=[{"type": "code_execution_20250825", "name": "code_execution"}],
)When a figure looks wrong, trace it from cheap checks to expensive ones, and change nothing until you know the cause:
- Read the method. Ask Claude to show the code and list the filters, definitions and date ranges it applied. The response already contains the commands it ran.
- Spot-check a sample. Compare the result for a small set of customers against the same rows in the original spreadsheet.
- Name the cause and correct the definition. For example, inactive accounts may have been counted as churned. Tell Claude exactly what churn means.
- Re-run with the fix. Run the analysis again with the corrected definition and confirm the new result against the same sample.
- Record it. Write the final figure, the definition used and the checks performed in the report's methods note.
This is iterative refinement: feeding Claude's output back in and asking it to verify or expand on what it produced. A re-run only helps once step 3 has changed something.
Checkpoint 5 of 8· Fill the gap
The team's sales CSV has to reach Claude's sandbox, which has no internet access. Which content block type references the uploaded file so that code execution can analyse it?
# Upload a file
file_object = client.files.upload(file=Path("data.csv"))
# Use the file_id with code execution
response = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "Analyze this CSV data"},
{"type": " ? ", "file_id": file_object.id},
],
}
],
tools=[{"type": "code_execution_20250825", "name": "code_execution"}],
)The code execution container cannot download anything, so the team's own data is uploaded through the Files API. The request then references it with a container_upload block, which puts the file inside the sandbox for Claude's code. A document block is used for citations and passes the file to the model as text to cite. It does not give Claude's code a file to compute on.
Source: platform.claude.comCheckpoint 6 of 8· Exam question
Claude's code analysis of a customer file reports a churn rate of 41%, but the team expects about 8%. An analyst must work out where the number came from before it goes into a report. Which sequence should the analyst follow?
Correct answer: A — Ask Claude to show its code and filters → Spot-check sample customers against raw spreadsheet rows → Pin down the cause and set the definition → Re-run analysis and compare result with sample → Record figure, definition and checks in methods note
- A. Reading Claude's method is the cheapest way to see where the figure came from. A spot check against raw rows then narrows down which definition is wrong. Only once the cause is known is the definition fixed, the analysis re-run and compared with the sample, and the outcome recorded.
- B. The spot check is placed before reading the code and filters. Without knowing the method, the analyst cannot tell which definition or filter a mismatch points to. The free look at the code should come first, because it often reveals the problem without any manual comparison.
- C. The cause is named and the definition given to Claude before the sample is checked. A cause guessed from the code alone has not been tested against the raw rows, so the analyst may correct the wrong filter or definition. The spot check is what narrows down the cause.
- D. The definition is set before anyone has looked at Claude's code and filters. Nothing yet shows which definition or filter produced 41%, so the instruction to Claude is a guess. The method has to be seen, then tested against the sample, before the cause can be named.
5.Synthesis into an actionable plan, then review
Synthesis is the stage where findings become decisions. Ask Claude to combine the verified research and the computed figures into a plan, and be specific about the output. If the brief calls for actions, owners, dates and risks, ask for exactly those, as numbered steps where order matters. Tell Claude to use only the material you provide and not its general knowledge. This keeps unverified claims from slipping back into the plan at the last step. A useful self-check: after drafting, Claude looks for a supporting quote for each claim and removes any claim it cannot support.
Nothing leaves the team until a person has reviewed it. The reviewer checks the plan against the brief written in stage 1 (audience, decision, limits, success criteria) and checks its figures against the analysis in stage 3. How deep that review should go depends on the stakes, which is covered in review depth and escalation. It sits last because only a finished plan can be compared with the brief as a whole.
The plan's actions should follow from verified findings and computed figures. If the plan is drafted first, the evidence ends up being chosen to fit conclusions already written, and the reviewer can no longer tell which claims the data actually supports.
Checkpoint 7 of 8· Check yourself
During synthesis, Claude's draft plan includes a market-growth figure that appears in neither the verified research nor the computed analysis. What instruction would have kept it out?
Telling Claude to use only the provided material stops unverified general-knowledge claims from getting back into the plan at the last step. The quote-per-claim self-check then removes anything still unsupported. Running a new search reopens research that has not been verified. A numbered format controls structure, not sources. Human review checks the plan against the brief and the analysis, and it does not replace grounding the synthesis.
“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”Source: platform.claude.com
Checkpoint 8 of 8· Exam question
A strategy team uses Claude to produce a market-entry plan. The sales analysis on the uploaded data is designed around the questions the research surfaces, and a team member must sign off before the plan leaves the team. Which sequence of stages should the team follow?
Correct answer: D — Write requirements brief for the plan → Run cited research and open key sources → Run code on the uploaded sales data → Draft plan with actions, owners, dates, risks → Team member checks plan against brief
- A. Research is run before the brief is written, so Claude has no stated audience, decision or scope to aim at. The sources gathered may not fit what the plan must support, and the research may need redoing once the brief exists.
- B. The sales analysis is run before the research. The analysis is designed around the questions the research surfaces, so computing figures first means the code may calculate the wrong things and have to be rerun.
- C. The brief is written only after the research and the analysis are done. Both stages then proceed without a defined audience, decision or scope, so the work is not aimed at what the plan must support. The brief also arrives too late to guide them.
- D. The brief sets the audience, decision and scope first, so research has a target. Research and its sources then surface the questions the sales analysis is built around. The plan is drafted from both results and reviewed by a person against the brief before it leaves the team.
Sources
(7)
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/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure”
↩︎ Why a deliverable is built in stages“Each step is a separate API call so you can log, evaluate, or branch at any point.”
↩︎ Why a deliverable is built in stages“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Requirements analysis: the brief comes before the draft“can give Claude more information about your goals and help it deliver more targeted responses”
↩︎ Requirements analysis: the brief comes before the draft“This structured approach helps Claude work through large corpora methodically and iteratively critique its findings.”
↩︎ Research with cited sources“Provide instructions as sequential steps using numbered lists or bullet points when the order or completeness of steps matters.”
↩︎ Synthesis into an actionable plan, then review“is still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure”
↩︎ Key concept“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Exam trap 1“Explicit prompt chaining (breaking a task into sequential API calls) is still useful when you need to inspect intermediate outputs”
↩︎ Checkpoint - 2.
“Specific: Clearly define what you want to achieve.”
↩︎ Requirements analysis: the brief comes before the draft“Measurable: Use quantitative metrics or well-defined qualitative scales.”
↩︎ Requirements analysis: the brief comes before the draft - 3.
“allowing it to answer questions with up-to-date information beyond its knowledge cutoff”
↩︎ Research with cited sources“Claude searches when the request depends on information that is current, changing, or outside its training data”
↩︎ Research with cited sources - 4.
“When people conduct research, they tend to continuously update their approach based on discoveries, following leads that emerge during investigation.”
↩︎ Research with cited sources - 5.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“by having it cite quotes and sources for each of its claims”
↩︎ Research with cited sources“asking it to verify or expand on previous statements”
↩︎ Data analysis through code execution“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Synthesis into an actionable plan, then review“For each claim, find a direct quote from the documents that supports it.”
↩︎ Synthesis into an actionable plan, then review“Make Claude’s response auditable by having it cite quotes and sources for each of its claims.”
↩︎ Checkpoint - 6.
“Check that each field has a value and at least one citation.”
↩︎ Research with cited sources“Set citations.enabled=true on each of your documents.”
↩︎ Research with cited sources“Check that each field has a value and at least one citation.”
↩︎ Exam trap 2“To get a citation for each extracted value, prompt Claude to answer with one line per field.”
↩︎ Prediction - 7.
“The container has no internet access”
↩︎ Data analysis through code execution“To analyze your own data files (such as CSV, Excel, or images), upload them through the Files API and reference them in your request.”
↩︎ Data analysis through code execution“If you want Claude to run code for a borderline request, ask explicitly”
↩︎ Data analysis through code execution“The response interleaves server_tool_use blocks (the commands Claude ran) with their tool result blocks”
↩︎ Exam trap 3“The response interleaves server_tool_use blocks (the commands Claude ran) with their tool result blocks”
↩︎ Prediction