# Vendor Due Diligence Vendor spend review for an acquired company, prepared for a $1B parent. It classifies every vendor as Terminate, Consolidate or Optimize, and estimates the savings. ## How this was done All the work was done in the Claude Code CLI, with me (Angelika, VP of Operations) directing it and approving each decision: 1. **Rules:** I wrote the project rules into `CLAUDE.md`, which Claude Code reads at the start of every session. 2. **Profile:** a script profiled the input and flagged non-vendor rows, garbled names and near-duplicates. 3. **Rubric:** Claude Code proposed the rubric and savings rates. I reviewed and approved them before any vendor was classified. 4. **Classification:** Claude Code classified the vendors in batches of 40 against the rubric, and a script checked each batch. 5. **Verification:** parallel Claude Code agents web-checked the largest vendors, a random sample and the unrecognized vendors. Every change is logged. 6. **Savings:** Python scripts calculated every figure from `outputs/assumptions.json`, never by hand. 7. **Deliverables and QA:** Claude Code drafted the Top 3, memo and methodology, and scripts checked each figure against the data. 8. **Final review:** three independent agents reviewed every row, every document and the repo from a clean copy. Their fixes are logged as `final_review` in `qa/corrections_log.csv`. Every prompt is saved in `prompts/`, two of them as marked summaries, and git history shows each stage. Project rules are in `CLAUDE.md`. The rubric and rates are in `outputs/rubric.md` and `outputs/assumptions.json`. ## Folder map | Folder | Contents | |---|---| | `inputs/` | The original vendor CSV, never changed | | `prompts/` | Every prompt given, in order | | `scripts/` | All calculations and checks | | `docs/` | The assessment brief | | `outputs/` | Deliverables: `memo.md`, `memo.docx`, `top3.md`, `methodology_tab.md`, `vendors_classified.csv`, `savings_model.md`, `savings_by_vendor.csv`, `rubric.md`, `assumptions.json`. Working files: `function_groups.txt`, `top3_tab.tsv`, `sheet_data.csv` | | `outputs/sheet_load/` | The blocks written to each tab of the Google Sheet | | `outputs/batches/` | Classified batches, which are the source for the merge | | `qa/` | Profile, lease sizing, validation, verification results, spot check, unrecognized-vendor web pass, corrections log, memo check, sheet check and open items | ## Run order All scripts use the Python standard library and never change `inputs/`. The one-page memo check also uses Chrome if it is installed, and is skipped otherwise. | Step | Script | Output | |---|---|---| | Profile the input | `scripts/01_profile.py` | `qa/01_profile_report.md` | | Merge the classified batches | `scripts/02_merge.py` | `outputs/vendors_classified.csv` | | Calculate savings | `scripts/03_savings_model.py` | `outputs/savings_model.md`, `outputs/savings_by_vendor.csv`, `outputs/sheet_data.csv` | | Validate the output | `scripts/04_validate.py` | `qa/04_validation.txt` | | Select the Top 3 | `scripts/05_top3.py` | `outputs/top3.md`, `outputs/top3_tab.tsv` | | Check every figure | `scripts/06_check_memo.py` | `qa/06_memo_check.txt` | | Build sheet data blocks | `scripts/07_sheet_load.py` | `outputs/sheet_load/` | | Build the Word memo | `scripts/08_memo_docx.py` | `outputs/memo.docx` | | Check the live sheet | `scripts/09_check_sheet.py `, using an anonymous export from the sheet's `/export?format=xlsx` link | `qa/07_sheet_check.txt` | These were one-time steps, so don't rerun them: `scripts/02_lease_and_top_vendors.py`, `scripts/03_verify_sets.py`, `scripts/04_apply_corrections.py`, `scripts/10_apply_unrecognized_pass.py` and `scripts/11_apply_final_review.py`. `03_verify_sets.py` and `04_apply_corrections.py` relied on the first-pass file, which has been replaced. Their results are in `qa/02_lease_and_top_vendors.md`, `qa/05_spot_check.md` and `qa/corrections_log.csv`. `outputs/memo.md` and `outputs/methodology_tab.md` are written by hand, so rerun the check after any change to the data. ## Google Sheet Links: [Google Sheet](https://docs.google.com/spreadsheets/d/1lwBvnjQGHllCJW62CdD8YosYNeBjFFrJVIqW6l-X9KM/view) and [memo](https://docs.google.com/document/d/1-HGvrALlU2IW_QA5PhUrHKgE-S3VA_mY6phxbT4i8t8/view). Claude Code wrote the Google Sheet directly through the Google Sheets connector, with no hand-pasting. It also created the memo as a Google Doc through the Drive connector and linked it from the CEO/CFO Recommendations tab, as that tab asks. Vendor names and costs in the sheet were left untouched. `scripts/07_sheet_load.py` builds the blocks written to each tab, in `outputs/sheet_load/`. The template's Department dropdown lacks four of our values, so Other, HR/People, IT/Security and Data/Analytics were added to its `Config` tab. `scripts/09_check_sheet.py` compares a public export of each tab with our files, row by row, and tests that the memo link opens. Results are in `qa/07_sheet_check.txt`. `outputs/vendors_classified.csv` stays the full working file, including Vendor_Name_Clean. ## Limitations - **Evidence:** the input has vendor names and 12 months of spend only, with no usage, contract, renewal or owner data. Names and costs were modified. - **Rates:** all savings rates and one-time costs are planning judgments, not benchmarks. - **Parent contracts:** rows labelled Consolidate into a parent contract assume that contract exists. This hasn't been confirmed. - **Verification coverage:** 124 of 386 rows were web-checked, covering 84.27% of spend. The rest are first-pass judgments, and sample accuracy is in `qa/05_spot_check.md`. - **Verification agent conduct:** one verification agent searched outside its instructions, using extra search words and fetching web pages directly. No spend or project data was sent. - **Late identifications:** 31 small vendors were identified after the savings model was approved. They were not assessed under the rubric and carry no saving (`qa/unrecognized_pass.csv`). - **Open data needs:** items for the business to confirm are in `qa/open_items.md`.