Most carbon footprint prompts circulating online ask for the same thing: "calculate my carbon footprint." That is exactly the task a language model fails at, because it has no factor database behind it. These forty prompts do the opposite. They hand the generative AI the scoping, structuring, sorting, and checking work, which makes up most of a real exercise, and leave the calculation to a tool that knows where its coefficients come from.
The order follows a real carbon footprint: scope, sources, collection, accounting export, GHG Protocol categories, emission factors, checks, reporting, reduction, verification. Each prompt is written out in full, with what to attach and what the model cannot do. Works with ChatGPT, Claude, and Gemini. For the broader set of prompts covering the whole CSR role, see our guide to AI prompts for CSR managers.
The warning to read before the first prompt
A language model produces plausible text. Nothing in how it works distinguishes a real emission factor from an invented one that looks real. Ask for the road transport factor in France and you will get a number, a unit, sometimes a full reference. That number might be correct. It might also be a statistical reconstruction of what a factor looks like, with no real data line behind it.
Three rules, non-negotiable:
- No number produced by a prompt enters a carbon footprint without being traced to an identified database, with its vintage and scope. If you cannot point to the line, the number does not exist.
- References need checking too. A vintage year, a standard clause, or a factor ID cited by a model is a lead to verify, not a source.
- Validation stays human. AI proposes structure and a first draft. A named owner validates the scope, the assumptions, and the result.
Verification checkpoints are covered in more depth in our guide to AI-assisted carbon accounting methods.
Three habits before you start: attach real data, because a prompt without a file produces a generic example. Anonymize, by stripping client names and sensitive amounts before sending anything into a public model. And close every prompt with "flag anything you're not sure about," which turns a confident answer into a checklist.
Step 1 · Scope the boundary
The boundary decides everything else. These prompts make it explicit before the first line of data.
1. Operational control vs. equity share
"Here are my entities with ownership stakes and governance details. Give me the boundary under operational control, then under equity share, and flag where they diverge."
Attach : entities, ownership stakes, governance.
Limit : a poorly described joint venture will be misclassified.
2. Draft the boundary memo
"Write a one-page memo for my steering committee: entities included, entities excluded, the reason for each exclusion, reporting year, gases covered, and decisions still to be made."
Attach : the list from prompt 1.
Limit : the memo reflects your choices, it does not validate them.
3. Spot the weak exclusions
"Here are my boundary exclusions and their justification. Play the skeptical reviewer: what question would you ask, what evidence would you demand? Rank from strongest to weakest."
Attach : the exclusions and their rationale.
Limit : this simulates an audit, it is not one.
4. Reconcile the carbon boundary with the financial boundary
"Compare the entities in the financial consolidation scope with those in the carbon boundary. Flag the gaps in both directions and the question to raise with finance."
Attach : both lists, a shared identifier.
Limit : text-based matching, a renamed entity will slip through.
Step 2 · Inventory your data sources
Knowing what already exists, and where. The most productive step for a model, because it is purely combinatorial.
5. Map sources by activity
"Here is my activity and my sites. List the internal sources that likely exist for a carbon footprint: document, owning department, format, frequency. No factors, no estimates."
Attach : sector, headcount, sites, activities.
Limit : a generic list until checked against reality on the ground.
6. Recover data already produced for another purpose
"List the regulatory, accounting, or contractual documents my sector already produces that contain carbon-relevant data. For each: what data to extract, what transformation to apply."
Attach : sector, country, known obligations.
Limit : the obligations cited vary by country, verify them.
7. Rate a source's reliability
"For each source, propose a 1-to-5 score based on its nature: metered, invoiced, self-reported, estimated. Say what would move it up a notch."
Attach : the sources and their nature.
Limit : the rating targets the source type, not the file itself.
8. Identify gaps before they become blockers
"Here is my boundary and my source inventory. Identify the categories with no source. For each, propose three options: substitute data, documented estimate, justified exclusion. Do not choose for me."
Attach : boundary and inventory.
Limit : the options are not methodologically ranked.
Step 3 · Prepare data collection
Response rates depend mostly on how clear the request is. A writing problem, so good ground for AI.
9. Draft the request to an internal team
"Write a 150-word email to facilities asking for this data for the given period. For each item: format, unit, acceptable source. Direct tone, no carbon jargon."
Attach : the data points, the period, the deadline.
Limit : the model doesn't know internal sensitivities, review the tone.
10. Build the data entry template
"Propose the structure of a collection spreadsheet for ten sites, one row per site and per data point: column names, required unit, date format, comment field, one example row."
Attach : the list of data points, the number of sites.
Limit : the structure must still import cleanly into your tool.
11. Write the entry instructions
"Based on this structure, write the instructions for contributors, one instruction per column. Specify what to do if a value is unknown, partial, or in a different unit."
Attach : the structure from prompt 10.
Limit : instructions don't replace checking data on receipt.
12. Draft the request to a supplier
"In 200 words, draft a request for emissions data to a supplier: metered data first, fallback to an average per unit sold. State that no response will lead to a default estimate."
Attach : nature of the purchase, annual volume.
Limit : the model doesn't know your commercial leverage.
Step 4 · Structure an accounting export
The general ledger export is the densest, most tedious source. These prompts handle the labels, never the amounts.
13. Read an unfamiliar chart of accounts
"Here are my accounts with description and annual total. Indicate which ones may contain emission-generating spend, and how. Rank by amount and flag descriptions that are too vague."
Attach : the account export, no personal data.
Limit : a description doesn't tell you what the account actually contains.
14. Clean up transaction labels
"Here are 200 raw transaction labels. Normalize them: fix abbreviations, strip invoice numbers and dates, group labels of the same nature together. Return a mapping from original to normalized label."
Attach : raw labels, anonymized.
Limit : grouping by text similarity, not by actual spend category.
15. Filter out lines with no physical flow
"In these accounting lines, identify the ones with no physical flow: provisions, depreciation, adjustments, internal transfers. Justify each exclusion in one sentence. Flag ambiguous cases instead of deciding."
Attach : account, label, amount.
Limit : group-level restatements often go unnoticed.
16. Spot double counting across sources
"Here are energy spend records from accounting and consumption records from invoices. Identify the sites and periods present in both, and which source to keep in each case."
Attach : both datasets, sites and periods aligned.
Limit : assumes consistent site identifiers.
Step 5 · Map flows to GHG Protocol categories
Classification with written rules. This is where AI helps most, especially for Scope 3 upstream and downstream categories, the main source of errors.
17. Decide between Scope 1 and Scope 2
"For each energy flow, state whether it falls under Scope 1 or Scope 2 per the GHG Protocol, justified by control of the energy source and by the location of the combustion."
Attach : energy source, use, location.
Limit : district heating and self-generation need manual review.
18. Assign a spend line to a Scope 3 category
"For each spend type, propose the most likely Scope 3 category per the GHG Protocol, with number and justification. If two categories could apply, give the deciding criterion."
Attach : the spend types from step 4.
Limit : errors cluster at category boundaries, check the largest items.
19. Document the boundary between two neighboring categories
"Explain the boundary between Scope 3 categories 1 (purchased goods and services) and 2 (capital goods), then between 4 (upstream transportation and distribution) and 9 (downstream transportation and distribution), with three examples per category and one counterexample that tips into the neighboring one."
Attach : your sector, for relevant examples.
Limit : cross-check against the standard's text before citing it.
20. Justify a non-applicable category
"Here are the fifteen Scope 3 categories and my activity. For each category that doesn't apply, write a two-sentence exclusion justification, usable as-is in a report."
Attach : activity, business model, physical flows.
Limit : size it before excluding, an excluded category can still be material.
Step 6 · Select and justify emission factors
The riskiest step for naive use of generative AI. The rule: the model describes what to search for, never the value. If the concept is new, start with our definition of an emission factor.
21. Describe the factor to search for, without the value
"For this flow, tell me what emission factor to search for: unit in the denominator, life-cycle scope, gases covered, geographic granularity. Give no numeric value or identifier."
Attach : the flow description and your unit.
Limit : none, as long as the restriction stays in the prompt.
22. Check unit consistency
"Here is my activity data with its unit, and the units expected in the denominator of the factors I've selected. Flag each mismatch and give the conversion formula, not the result."
Attach : both lists, explicit units.
Limit : density and calorific value need separate documentation.
23. Write the justification for a factor choice
"I selected this factor, from this database, for this vintage year. Write a three-sentence justification for the methodology appendix: scope fit, geographic fit, temporal fit. Don't downplay the limitations."
Attach : the exact reference, retrieved by you from the database.
Limit : it writes up your choice, it doesn't validate it.
24. Compare two candidate factors
"Here are two candidate factors for the same flow, with scope, vintage, region, and uncertainty. Compare them, recommend which to use, and say what would flip the recommendation."
Attach : both factor sheets copied from the databases.
Limit : it compares what you give it, without searching for better options.
Step 7 · Check consistency and flag anomalies
Comparison work, where a model spots things a spreadsheet won't flag on its own.
25. Detect outliers in a data collection
"Here is the data collected by site, with headcount and floor area. Calculate ratios per employee and per square meter, isolate outliers from the median, and propose three possible explanations."
Attach : the table and the normalization data.
Limit : an outlier can be real, never correct it mechanically.
26. Compare two years line by line
"Here is my activity data for the last two years, at the same level of detail. Identify variations above 20%, categories that appeared or disappeared, and the question to raise with the contributor."
Attach : two datasets with matching structure.
Limit : a boundary change produces false positives.
27. Check the shape of the result
"Here is my emissions breakdown by category, in percentages. Flag any category whose share looks unusual for my sector, phrasing each flag as a question to verify."
Attach : the percentage breakdown and the sector.
Limit : these benchmarks are indicative only, never cite them as a reference.
28. Write the pre-closing checklist
"Based on my method and my sources, write the checklist to run before locking the results: one check per line, from most to least critical."
Attach : your method and your sources.
Limit : a generic list, to be completed with on-the-ground input.
Step 8 · Write the report and the assumptions
The only ground where AI works with no risk of fabricating a number, as long as you feed it only your own figures.
29. Write up the assumptions log
"Here are my assumptions in raw notes. Structure them: assumption, rationale, estimated effect on the result, condition that would invalidate it. Keep my figures exactly as given, don't add or round anything."
Attach : your notes, with figures already calculated.
Limit : check that no number was rephrased in the process.
30. Explain a result to senior leadership
"Here are my results by scope and by category. Write a one-page summary for a non-technical committee: what the number covers, what it doesn't, the three biggest categories."
Attach : consolidated results and uncertainty.
Limit : it reuses your figures, it doesn't recalculate them.
31. Write the limitations section
"Based on these known weaknesses, write the limitations and uncertainty section. Factual tone, no hedging language. End with what will be improved next reporting cycle."
Attach : the weaknesses, including those from prompt 8.
Limit : the model tends to soften language, demand no hedging.
32. Adapt one piece of content for three audiences
"Here is my summary. Produce three versions: executive committee at 200 words, internal teams at 150 words, requesting client at 100 words. No difference in figures across the three."
Attach : the validated summary.
Limit : check that all three carry the exact same numbers.
Step 9 · Build the reduction plan
The plan is built from your own categories, not from a generic catalog of actions.
33. Trace back to physical levers
"Here are my emissions by category and the underlying activity data. For the five heaviest categories, identify the physical lever at play: volume, intensity, mix, or usage rate."
Attach : detailed results and activity data.
Limit : the lever may be outside your control.
34. Build out a single action in depth
"Detail this action: reduction mechanism, data needed to size the effect, realistic timeline, internal prerequisites, three reasons it could fail. Don't quantify the gain, explain how to quantify it."
Attach : the action and the site's context.
Limit : quantification needs your own data and factors.
35. Prepare the trade-off between several actions
"Here are five actions with estimated gain, cost, and timeline. Build the decision matrix: criteria beyond cost per tonne, dependencies between actions, launch order, and which one to start first."
Attach : the five actions, costed by you.
Limit : cash flow and contractual deadlines are unknown to it.
36. Anticipate internal objections
"Here is my reduction plan. For each action, phrase the most likely objection from finance, operations, and the affected teams, then the factual point that would answer it."
Attach : the plan and your organization.
Limit : plausible objections, not an exhaustive list.
Step 10 · Respond to a verifier
Verification is about traceability, not the result. These prompts prepare the defense of the file.
37. Simulate the review before it happens
"Here is my methodology note and my assumptions. Play the verifier: the twenty questions you'd ask first, ranked by risk, with the supporting document requested for each. Don't write the answers."
Attach : the full note and the assumptions.
Limit : an incomplete note will produce easy questions.
38. Reconstruct the audit trail of a number
"Here is a category from my carbon footprint and, in random order, the files, transformations, and factors used to calculate it. Reconstruct the chain and flag every break in traceability."
Attach : the steps and files, even if incomplete.
Limit : it sees what you describe, not what you left out.
39. Draft the response to a justification request
"Here is the verifier's question and my supporting elements. Write a short response: what was done, on what basis, with what evidence. If it's not enough, list what's missing instead of filling the gap."
Attach : the exact question and your evidence.
Limit : don't drop the instruction not to fill the gap.
40. Prepare next year from the review comments
"Here are the review comments. Sort them into three groups: data, method, documentation. For each, what needs to be set up before the next collection cycle, with a typical owner and a deadline."
Attach : the comments, as given.
Limit : helps prioritize, it's not a validated plan.
What Kabaun changes about these prompts
These forty prompts share one constraint: the model must never produce an emission factor from memory. That's the problem Kabaun's architecture solves.
Klem, the assistant built into the platform, answers in natural language on your data, your results, and applicable regulations, directly inside the interface. It detects anomalies on import, categorizes accounting lines, extracts data from PDF invoices, and generates prioritized action plans with an estimated CO2e gain.
The native MCP architecture exposes API tools that fetch data from an identified database instead of answering from memory. The full ADEME Base Carbone is accessible through it. An external assistant connects through delegated authentication: it acts within your perimeter, with your permissions, and every action is logged. With Klem, the principle stays: Klem proposes, you validate.
Frequently asked questions
Can you calculate a carbon footprint directly with ChatGPT?
No. A language model has no queryable factor database and will produce a plausible but unverifiable coefficient. It is effective, though, at scoping the boundary, structuring data collection, cleaning accounting labels, mapping flows to GHG Protocol categories, and drafting the report. The calculation itself needs a tool that traces the origin of every factor.
How do you stop a model from inventing an emission factor?
By explicitly forbidding it from producing a numeric value. Frame the request as a research brief: what unit in the denominator, what life-cycle scope, what gases, what geographic zone. You then look up the value yourself in an identified database. That's the principle behind prompt 21.
Can these prompts be used with confidential data?
Not in a public model without precautions. Strip client names, personal data, and sensitive amounts before sending anything. Most of these prompts work fine on anonymized labels and data structures, without real values. Otherwise, require written guarantees from the tool you use on hosting and on how your data is used.
Which model gives the best results?
All forty prompts work with ChatGPT, Claude, and Gemini. Long tasks like file structuring and methodology writing benefit from models with large context windows, able to process a full accounting export in one pass. No model removes the need to verify the numbers.
Can a carbon footprint prepared with AI be verified by a third party?
Yes, if the audit trail is intact. The verifier doesn't check the drafting tool: they check the origin of the data, the justification of the factors, the documentation of the assumptions, and the traceability of the calculations. A carbon footprint prepared with AI is verified like any other, as long as every number is traced back to an identified source.
From a list of prompts to a traceable exercise
These forty prompts save time on everything that surrounds the calculation. They replace neither a factor database nor the validation of a named owner. The real question is chaining: how many back-and-forths between a model, a spreadsheet, and a factor database are you willing to accept, and where does traceability break.
Let's talk about your carbon footprint exercise and what can be automated without losing the audit trail.
