Report Section Draft and Claims Check: A Four-Prompt Claude Chain That Drafts From Approved Inputs and Audits Every Sentence

What This Builds

Sustainability report sections tend to fail in the same two ways. A figure gets rounded or restated differently from the source, or a sentence promises more than the evidence shows. This chain tackles both by splitting the job into four prompts in one Claude conversation. The first checks your inputs, the second drafts the section using only those inputs, the third audits every sentence against them, and the fourth revises.

You finish with a draft in which every figure and factual sentence carries an input ID, a table showing which sentences have no input behind them, and a short list of what you still have to check. Claude drafts, sorts and flags. It never decides that a claim is substantiated, and it never produces an official number. Your spreadsheet does that.

Prerequisites

• A Claude account. The chain runs in a single conversation, and a long one uses more of your plan's usage allowance, so check your plan on claude.ai. The entry paid plan is Pro ($20/month at the entry level), and that is the total ongoing cost of this build.
• A set of approved inputs for one report section: figures with their units and periods, statements already approved for publication and notes on the calculation method, all written down in your own words and numbers
• The source data behind each figure, open in a spreadsheet beside you
• About an hour for a first run

Before you paste anything. Report text and figures that have not been published can be material non-public information for a listed company, and a consultant's engagement letter or NDA decides what client material a third-party tool may see. Unreleased report text goes only into a tool your company has approved for it. Open the privacy settings of your Claude account and check whether chats are used to train models and how long they are kept, because the answer depends on the plan. Replace the company or client name with a project code or "Company A", leave out facility addresses, account numbers and the names of employees or supplier contacts, and use aggregated figures rather than individual-level data.

The Concept

Think of a careful editor who will only write down what you hand across the table. Each item on the table gets a label. The editor reads the items back to you first, asking about anything that is missing or does not fit. Then the editor drafts, tagging each sentence with the label of the item behind it. Then the editor goes through the draft again, this time looking for anything that has no label or sounds stronger than its label allows.

A chain works this way because each prompt does one job and you check its output before the next one starts. If the model's first reading of your inputs is wrong, you want to find out before it writes four paragraphs on top of it.

Build It Step by Step

Part 1: Prepare the Inputs

Number each input with an ID: I1, I2, I3 and so on. Write each on its own line with its unit and the period it covers. A figure line looks like "I2: Total waste generated, 2025: 1,200 tonnes, calendar year, all sites." An approved statement line looks like "I4: Policy statement, approved text: [paste the approved wording]." A method line says how the figure was calculated and where the factor or method came from, in your words.

Calculations stay in your spreadsheet. If a section needs an emissions figure, you calculate it there with a published emission factor you name, and you paste the result in as an input. Do not ask Claude for a factor or a total.

Start a new Claude conversation, ideally inside a Claude Project if you use one for this report so your approved style notes are on hand.

Part 2: Prompt 1, Inputs Check

Paste this, followed by your inputs.

I am drafting the [section name] section of a sustainability report for Company A. Below are the approved inputs, each with an ID. Do not draft anything yet.

Do these things:
1. List every input back to me in a table with its ID, what it measures, its value, its unit and its period.
2. Flag any input that is missing a unit or a period.
3. Flag any inconsistency between inputs, such as different boundaries, periods or units that would make a comparison unreliable.
4. List any figure that someone might reasonably derive from these inputs, such as a change or a share, and say which inputs it would use. Do not calculate or report these figures.
5. Ask me every question you need answered before a draft can use only these inputs.

Inputs:
[paste inputs here]

Checkpoint 1. Compare the table with your inputs line by line. Look for a value that changed on the way back, a unit that was filled in by the model, and a period that was assumed. Answer every question in the same conversation. If an answer changes an input, restate the input with its ID. Move on only when the table matches your source and the questions are settled.

Part 3: Prompt 2, Draft

Draft the [section name] section for the report, in about [length] words, in a plain factual tone.

Rules:
- Use only the inputs we have confirmed. Add nothing from your own knowledge.
- After every figure and every factual sentence, add the input ID in square brackets, for example [I2].
- Where the section needs information we do not have, write a placeholder in square brackets that says exactly what is missing, for example [add the diversion method].
- Do not calculate new figures. If a change or share is needed, write a placeholder naming the inputs it would use, and I will calculate it in my spreadsheet.
- No promotional wording and no claims about the company's character or standing.

Checkpoint 2. Read the draft against the inputs. Every figure must match its input exactly. Every sentence with a fact has an ID. Placeholders mark real gaps. Now do the part Claude cannot: calculate any change or share in your spreadsheet and put the result into the draft yourself, with the inputs it came from. This is also the point where you may edit the text, and colleagues' suggested wording often arrives here. Keep any edit you make, because prompt 3 audits the draft as it stands.

Part 4: Prompt 3, Claims Audit

Audit the draft below, exactly as it stands, including any wording I added myself.

Produce a table with one row for every sentence that states a figure, a fact or a claim. Columns:
- Number
- The sentence
- Type (figure, fact or claim)
- Input ID, or "no input" if no confirmed input supports it
- Flagged wording, if any
- Why it needs evidence, and what kind

Treat these words as flags whenever they appear: significant or significantly, leading, net zero, carbon neutral, zero waste, eco-friendly, sustainable, green, and any similar wording that goes beyond what the inputs show.

Also check that every figure in the draft matches its input exactly, and list any figure that does not. After the table, list any sentence that sounds stronger than its input supports, even if no flagged word is used. Do not rewrite anything yet. Do not say that any claim is substantiated.

Draft:
[paste the draft as it now stands]

Checkpoint 3. Read every "no input" row and every flagged row. Add any sentence the table missed. Ask for another pass if the table skipped a paragraph. Confirm each reason is something you agree with. Decide which unsupported sentences you want fixed and which you want removed.

Part 5: Prompt 4, Revise

Revise the draft using the audit table.

Rules:
- For every sentence marked "no input", either turn it into a question for me or rewrite it with more cautious wording that stays within the inputs. Do not invent support for it.
- Replace flagged words with plain description of what the inputs show, or remove the sentence.
- Keep the input IDs in square brackets so I can review them.
- Keep placeholders for anything missing.
- At the end, give me a short checklist of what I must check myself before this goes anywhere: each figure against the source data, each claim with legal counsel, and the assurance provider if the company uses one.

Checkpoint 4. Compare the revised draft with the audit table, sentence by sentence. Check each figure against your source data, not against the inputs list. Send every environmental claim to counsel. If the company uses an assurance provider, check what it needs to see. Remove the input IDs only when you move the text into the final document, and keep a copy with the IDs for the reviewers.

Part 6: Reuse It

Save the four prompts in a document. For the next section, start a new conversation, change the section name and paste the new inputs. If a section spans several periods, number inputs by period so the audit can catch mixed periods.

Real Example: A Waste Section for Company A

Invented example, with invented figures and wording. Company A is a mid-sized manufacturer.

Inputs the reader pastes into prompt 1:

• I1: Total waste generated, 2024: 1,500 tonnes, calendar year, all sites
• I2: Total waste generated, 2025: 1,200 tonnes, calendar year, all sites
• I3: Waste diverted from landfill, 2025: 780 tonnes, calendar year, all sites
• I4: Policy statement, approved text: "Company A's waste policy commits the company to reduce waste sent to landfill and to report diversion each year."

Prompt 1 result. A typical response lists the four inputs and notes that I1 and I2 share a boundary and unit. It points out that I3 covers only 2025, so there is no 2024 diversion figure to compare with. It lists two derivable figures, the change in total waste (I1 and I2) and the diversion share (I3 and I2). It asks one question: what does "diverted from landfill" include, for example recycling, reuse and composting? The reader answers from the waste contractor's reporting notes and restates I3 with that wording.

The reader's check, by hand. The change in total waste is (1,500 - 1,200) / 1,500 = 300 / 1,500, a 20 percent decrease. The diversion share is 780 / 1,200, a 65 percent diversion rate. The reader calculates both in the spreadsheet, confirms that both match, and adds them to the draft after prompt 2 with their input IDs.

Prompt 2 draft (typical, abridged). "Total waste fell from 1,500 tonnes in 2024 to 1,200 tonnes in 2025 [I1, I2], a decrease of [reader to add the change, from I1 and I2]. In 2025, 780 tonnes were diverted from landfill [I3]. The company's waste policy commits it to reduce waste sent to landfill and to report diversion each year [I4]. [Add the diversion method.]"

Between prompts 2 and 3. The reader fills in the two placeholders with the 20 percent and 65 percent figures, then pastes in the communications team's headline suggestion at the top: "Company A reduced waste significantly while becoming a zero-waste company."

Prompt 3 audit, typical rows.

If the model's draft says anything like "420 tonnes went to landfill", the audit should also flag it as "no input". Subtracting 780 from 1,200 gives 420 tonnes that were not diverted, but the inputs do not say where those tonnes went, so landfill is an assumption. Catching that kind of inferred sentence is the reason for the audit.

Prompt 4 result. Sentence 1 becomes a question to the reader: what evidence supports "zero-waste", and what definition is used? If none, it is removed. A cautious replacement stays within the inputs: "Total waste fell by 20 percent between 2024 and 2025 [I1, I2], and 65 percent of 2025 waste was diverted from landfill [I2, I3]." The checklist lists the five figures to check against source data, the claim wording to take to counsel, and the assurance provider, if there is one.

Time saved: Claude does the first pass of cross-checking sentences against inputs in minutes. Your own time shifts from hunting for unsupported wording to checking the flagged items, and the checking against source data stays with you.

What to Do When It Breaks

• A figure in the draft does not match your input. Claude may have rounded or restated it. Paste the input back, tell Claude which sentence is wrong, and ask for that sentence only. Check the whole draft again for the same slip.
• The audit table skips sentences. Ask for a second pass over a named paragraph. Count the sentences yourself and compare with the table.
• Claude calculates a figure you did not ask for. Delete it. Calculations belong in your spreadsheet, and the draft should carry a placeholder.
• The conversation gets long and Claude loses an early input. Start a new conversation, paste the confirmed inputs and the current draft, and run prompts 3 and 4 again.
• Claude says a claim is fine. It cannot know. Substantiation is decided by you, counsel and, where one is used, the assurance provider.
• The model adds wording from its own knowledge. The input IDs make this visible: a sentence with no ID is the giveaway. Send it back to prompt 4.

Variations

• Simpler version: Run only prompts 2 and 3 on a short section where you trust the inputs.
• Extended version: Run the chain once per section, then paste all the revised sections into a final conversation and ask for a consistency check on figures and terms across the sections. Check the answers yourself.

What to Do Next

• This week: Run the chain on one small section with real inputs, in an approved tool.
• This month: Build a standing input sheet with an ID for every figure the company publishes, so inputs are ready when a report cycle starts.
• Advanced: Use the same inputs and audit prompts on questionnaire answers, so the report and the answers cannot drift apart.

Advanced guide for sustainability consultant and ESG analyst professionals. The chain drafts and flags only. Figures, claims and rule applicability are checked by you, counsel and any assurance provider.