supplier-compare

---
name: supplier-compare
description: Compares supplier quotes on a weighted 1-5 matrix and names a winner, so the decision is value not headline price. Use when quotes have landed and somebody must choose.
---

# Compare three quotes properly in ten minutes

A like-for-like comparison, a weighted score per supplier, a named winner, and the questions to ask before signing.

## Claude does

1. Ask for the quotes, what is being bought, and the purchase type: commodity, specialised component, capital equipment, professional services or logistics.
2. Choose 5 to 10 criteria. Fewer than 5 hides the trade-off, more than 10 buries it.
3. Set weights summing to exactly 100%, from the template for that type. Commodity: price 40-50, delivery 20-25, stability 15, service 10-15. Specialised components: quality 30-35, technical 20, price 20-25. Capital equipment: total cost of ownership 35-40, technical fit 25, support 20. Professional services: capability 40, price 25, cultural fit 20, references 15. Logistics: reliability 30-35, price 25, coverage 20. Show them; the user may change them.
4. Score every supplier on every criterion 1-5: 5 significantly exceeds, 4 exceeds, 3 meets, 2 partially meets, 1 fails.
5. Quote the line of the supplier's own document behind each score. No quotable line means "no evidence", never 3.
6. Normalise price first: same quantity, term, delivery, inclusions. No source publishes a normalisation formula, so print every assumption; silence is the failure here.
7. Apply the knockout: below 2 on quality capability or financial stability eliminates that supplier whatever the total says.
8. Refuse to guess. Missing lead times, payment terms, references and unpriced extras go on an "ask the supplier" list, not the matrix.

## Then Codex does

Codex takes the arithmetic and the paperwork: score times weight across every supplier and criterion, the ranking, the knockout, then the matrix, recommendation and question list as files. Bulk multiplication is where people slip.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in first: {PURCHASE_ITEM}, {CRITERIA_AND_WEIGHTS}, {SCORES_WITH_EVIDENCE}, {NORMALISATION_ASSUMPTIONS}, {OPEN_QUESTIONS}.

## Claude checks

1. Weights sum to 100.0.
2. Recompute one supplier's total by hand; match Codex to two decimals.
3. Prove the knockout fires: a supplier scoring 1 on stability is eliminated even when it holds the top total. A rule that never eliminates anybody is decoration.
4. Every score carries its evidence line, or reads "no evidence".
5. Prices cover the same scope and term; assumptions are on the page.

Any failure: reject and rerun, never patch by hand.

## Rules

- Weights sum to 100; never add a criterion nobody can score from the documents.
- Never compare headline prices across different scopes or terms.
- The knockout beats the total score, always.
- Two quotes is not a comparison; ask for a third (house rule).
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude does that half itself, and says so.

## Built from

- Anthropic, Skill authoring best practices, https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - the format spec obeyed here.
- AuraVMS, Weighted Quote Scoring Matrix, https://www.auravms.com/blogs/weighted-quote-scoring-matrix-supplier-evaluation - source of the criteria range, weights, anchors, templates and knockout rule.
- w95/awesome-claude-corporate-skills, https://github.com/w95/awesome-claude-corporate-skills - 189 stars from api.github.com; carries vendor-evaluation and supplier-scorecard.

# Compare three quotes properly in ten minutes - Codex task

You are scoring supplier quotes somebody else has already read. You have no memory of that conversation and no access to the original quotes. Work only from what is below. Do not open the internet.

## You are given

- {PURCHASE_ITEM}: what is being bought, and the purchase type (commodity, specialised component, capital equipment, professional services or logistics).
- {CRITERIA_AND_WEIGHTS}: 5 to 10 criteria, each with a percentage weight, summing to 100.
- {SCORES_WITH_EVIDENCE}: per supplier, a 1-5 score per criterion plus the quoted line from that supplier's document justifying it. Some entries read "no evidence".
- {NORMALISATION_ASSUMPTIONS}: the assumptions made to put prices on the same basis (quantity, term, delivery, inclusions).
- {OPEN_QUESTIONS}: what nobody could answer from the quotes.

## Produce

Three files in the working folder.

1. `matrix.csv` - one row per criterion; columns: criterion, weight, then per supplier a score column and a weighted column (score x weight / 100). Final rows: WEIGHTED TOTAL per supplier, and KNOCKOUT reading ELIMINATED or CLEAR.
2. `recommendation.md` - one page maximum, in this order: the recommendation in one sentence naming one supplier; a ranked table with weighted totals to two decimals; the two or three criteria that decided it; every "no evidence" entry listed as a gap; the normalisation assumptions verbatim; the open questions numbered, ready to send back to suppliers.
3. `workings.txt` - every multiplication as `criterion | supplier | score x weight = weighted`, then each supplier's sum, so the arithmetic can be checked line by line.

## Rules

- Compute, do not judge. Never change a score, add a criterion, or fill a gap.
- Treat "no evidence" as 0 in the arithmetic and label it in the output. Never substitute 3.
- Knockout: any supplier below 2 on quality capability or financial stability is ELIMINATED regardless of total. Show its total anyway, then mark it eliminated. Apply the knockout after the totals, never instead of them.
- If the weights do not sum to 100, stop, print `WEIGHTS DO NOT SUM TO 100` with the actual sum, and write no files.
- If two suppliers tie to two decimals, say so. Do not break the tie.
- Round only at display time, to two decimals.
- Copy assumptions and quoted evidence lines verbatim. Never paraphrase.
- Public information only. Invent nothing.
- Anything drafted for sending names whose name it goes out in.

## Return

Print, in order: the weights sum; the ranked suppliers with weighted totals to two decimals; any ELIMINATED supplier and the criterion that eliminated it; the count of "no evidence" entries; and the three file paths written.