Building a GHG Inventory with Claude — Instructor Demonstration
By the end of this session you can
- Set up a grounded Claude Project so the model works from your standards and your factors, not its memory
- Move messy activity data from raw workbook to classified, calculated inventory with an auditable trail
- Force calculation into executed code rather than generated prose, and verify the result
- Recognise the four failure modes — invented factors, silent arithmetic, plausible mis-mapping, confident gap-filling — and catch each one
Three-hour agenda
| 00:00–00:20 | Rules of engagement | What Claude is genuinely good at here — structuring, mapping, cross-checking, drafting, explaining — and what it must never do: supply a factor from memory, do arithmetic in prose, or be the sign-off. What leaves the building and what does not. |
| 00:20–00:45 | Grounding the workspace | A Project with custom instructions encoding the boundary, the period, the GWP set and the refusal rule; project knowledge loaded with the frozen factor library, the boundary memo, the site list and Protocol extracts. The same prompt with and without grounding, side by side. |
| 00:45–01:20 | Live build 1 — data triage | Upload the messy workbook. Profile it: what is here, what units, which months are missing, what is duplicated. Generate a plant-by-plant data request. Map the 4,000-line ledger to scope and category with a confidence column and a mandatory "needs human review" flag. |
| 01:20–01:30 | Break | |
| 01:30–02:15 | Live build 2 — calculation | Make Claude write and run code rather than compute in text. Constrain it to the supplied factor table, with an instruction to stop and ask if a factor is missing. Inventory with per-line factor citation; both Scope 2 methods with the PPA and solar treated correctly. |
| 02:15–02:40 | Live build 3 — outputs | From one dataset: inventory summary, per-category data-quality scorecard, chart pack, and an assurance-ready calculation trail where every figure traces to a data cell and a factor row. |
| 02:40–03:00 | Deliberate failure | The instructor prompts badly on purpose and produces a plausible, wrong answer. The class runs the verification protocol: recompute three lines by hand, check provenance, reconcile totals, find the silently filled gap. Then compare against Day 2's recorded times. |
Claude capabilities taught
- Projects: custom instructions and project knowledge as the grounding layer
- Spreadsheet upload and profiling before processing
- Forcing computation into executed code instead of generated text
- Output contracts: columns, flags and citation requirements specified up front
- Refusal instructions: "if the factor is not in the supplied table, stop and ask"
- Artifacts for summaries, charts and reusable calculation workbooks
Exercise
Class-wide verification of the instructor's deliberately flawed output. Every participant must find at least one defect and name which failure mode it belongs to.
Take-home
Individual company briefs and raw data packs issued for the Day 4 lab. Each participant gets a different sector: food processing, IT services, cement, pharmaceuticals, logistics, hospitality, chemicals, retail.