Skip to content
VIPS HandbookVIPS Handbook

AI-assisted editing, verification, and human review

Copy page

This handbook relies on AI assistance heavily, but final acceptance and review are owned by human maintainers. We state the method and boundaries transparently so that AI output is not mistaken for sourced fact.

The role of AI

Section titled “The role of AI”
  • Writing and translation: drafting Chinese and English content, keeping the bilingual structure and contentId aligned.
  • Code generation and execution: producing runnable examples (such as Go/govips prototypes) and actually executing them via scripts to produce evidence.
  • Cross-review: using scripts to check bilingual consistency, broken links, schema validation, and drift in registered fixtures or generated output.

How verification works

Section titled “How verification works”
  • Reproducible commands: before and after delivery we run checks such as pnpm check, pnpm build, and bilingual-consistency scripts to confirm the site builds and validates.
  • Script-generated evidence: images and metadata are produced and verified by a runner, never retouched by hand to fake production output.
  • Honest status: pages carry a reviewStatus (draft / in-review / approved); unverified conclusions stay draft or comparison-pending.

Version-sensitive claims

Section titled “Version-sensitive claims”

Any assertion that involves a specific version (for example govips v2.19.0, libvips 8.17.3) must name the verified version and the review date. AI generation does not automatically generalise “it runs on this version” to every version.

Where human review decides

Section titled “Where human review decides”
  • Only human maintainers may set reviewStatus to approved.
  • High-risk conclusions about colour, metadata, or formats must be checked against sources and evidence by a maintainer before adoption.
  • Third-party exceptions need their source, license, and attribution recorded separately; AI cannot decide ownership on its own.

Boundaries and limits

Section titled “Boundaries and limits”
  • AI assistance does not change the rule that conclusions need official sources; official docs, source code, and release notes come first.
  • Cross-binding semantic differences (govips vs sharp/pyvips/CLI) are only asserted after being genuinely compared by the shared runner, not guessed.
  • This project explicitly does not ship an AI-generated online Playground, nor does it treat that as equivalent to human review.