Review and comparison skill for your AI agent
Know what to improve first.
Review and comparison skill for your AI agent
Know what to improve first.
Give your agent a tool, page, or portfolio. CraftAssay checks whether its value comes across, weighs the evidence, and ranks useful improvements. A later review shows what changed.
Four review files, seven dimensions. An assessment, not a certification. The source stays unchanged.
HarborNote is a fictional notes app. These excerpts connect its sample review, the author’s revised page, and the later comparison.
1 · Original page Can my notes leave this device?
“Nothing leaves your machine. Ever.”
“Syncs to every laptop you own.”
The page offers both claims without explaining how they fit together.
2 · CraftAssay finding Fix the claim before polishing the pitch
harbornote-F001 · P1 · claim-risk
Surface: the page’s privacy and sync claims.
A privacy-seeking reader cannot tell what stays local. Name which data stays on the machine when sync is off.
P1 is the highest priority. The finding gives the author a specific correction to make.
3 · Author edit + later review The choice is now explicit
Notes stay on this device unless you turn on sync.
Later comparison: harbornote-F001 → resolved
Readers can distinguish local notes from optional sync. The review checks that the contradiction is gone.
This is a review of fictional copy, not a privacy audit. The author makes the edit. Scores support the finding: Accuracy moves from 3 to 6; Utility stays 6.
What it reads, writes, and changes
| Reads | A tool, site, skill, spec, portfolio, URL, or paste |
| Writes | A dated run folder: report.md, a scorecard, findings, and coverage |
| Changes | Nothing. Read the findings, then make the smallest useful correction yourself or ask the agent for that named edit |
| Label | Means |
|---|---|
| Review | Inspect, assess, score, and recommend |
| Findings | Specific issues with evidence and the smallest useful correction |
| Scores | Seven dimensions from 1 to 10, or NR (not rated) when evidence is missing |
| Evidence | What was inspected, and what was not |
| Compare versions | Finding status and explained score changes |
Accuracy is freedom from misleading claims in the inspected presentation. It is not a measured bug rate.
CraftAssay and Cold-eye
Cold-eye asks whether a newcomer can use what is presented as ready. CraftAssay asks how useful, clear, credible, and well-presented the work is. First-use friction still belongs in a CraftAssay finding when it affects those scores. CraftAssay does not issue a release-readiness verdict. A clean Cold-eye review does not mean CraftAssay will score the work highly.
Smell Check reviews unearned language. Detangler reviews what editing tangled.
Built by Catalyst Forge LLC. MIT. About