The reference skill for commonspecs — an API-first database of engineering-grade product specifications with per-field confidence. Install it, and your AI agent can look up what a product is actually made of (fabric weight and weave, shoe construction, country of origin), compare products on hard specs, and contribute specs it has verified.
npx skills add commonspecs/skillsWorks with Claude Code, Cursor, and the other agents the skills CLI
supports. Or install manually:
# Claude Code
cp -r skills/commonspecs ~/.claude/skills/Then connect. The best path is the commonspecs MCP server: it signs in with OAuth — no token to create or manage — and the skill drives its tools directly. If your agent has no MCP support, set an API token instead (get one at commonspecs.com):
export COMMONSPECS_API_TOKEN="cs_live_…"That's the only local configuration either way. Your buying preferences — market (country), quality and
locality strategy, contribution mode — live on your account
(commonspecs.com/account); the server applies them to every
read and returns them in a context block, so the agent never needs them configured locally.
Most product "data" online is marketing copy: adjectives, not specifications. When you ask an LLM "is this a good pair of raw denim jeans?", it pattern-matches vibes. commonspecs answers from hard facts — and tells you how confident each fact is and whether sources agree.
The point isn't a number to obey. It's to give your agent the facts that actually differentiate good from bad in a category, so it can reason about fitness for purpose and cost of use, not just price.
The skill teaches your agent to weigh the specs that matter (the exact scoring is intentionally not exposed):
- Raw denim / jeans — fabric weight (oz) and weave (selvedge vs open-end), composition, country/mill of origin.
- Welted footwear — construction (Goodyear-welted vs Blake vs cemented), leather grade, sole. Construction decides whether a shoe can be resoled — its real cost over years.
- Fragrance — concentration (parfum → EDP → EDT → EDC), longevity, projection.
The agent fetches the facts; it explains the trade-off in plain language. It never claims a value it didn't get back, and it always carries the confidence with the fact.
| Signal | Meaning |
|---|---|
quality_score (0–100) |
Overall spec quality; missing_fields shows what's dragging it down. |
confidence (0–1) |
How well-evidenced a value is. Always reported with the fact. |
fields[] |
Confidence ≥ 0.4 — solid enough to state. |
low_confidence_fields[] |
Confidence 0.1–0.4 — a lead, shown only with a hedge. |
needs_corroboration |
Only one independent source so far. |
disputed + alternate_claims |
Sources disagree; both values are shown. |
If your agent reads a verified spec (off a product page, or a physical label) that commonspecs is missing, it can submit it with the source URL and the exact snippet it read the value from — and a dated price observation from the same page, since specs and prices usually sit on the same fetch. Evidence is what earns confidence; corroboration from independent users is what makes a value trustworthy. Photos never leave your machine — only the extracted values and the product identity are sent.
See skills/commonspecs/SKILL.md for the exact API calls,
or the docs for the full API reference and methodology.
MIT © commonspecs