| name | package-review-prompt |
| description | Assemble a grounded review prompt from a local artifact bundle so Codex or Claude Code can inspect the right files and ask the right questions. |
Package Review Prompt
Use this when you want a reviewer agent to sound grounded in the real
run artifacts rather than hallucinating the review surface.
This skill produces:
- bundle summary
- artifact-aware prompt text
- explicit review focus areas
Inputs
Use the same bundle locators as video-artifact-intake, plus optional review
context:
{
"latestPointerRoot": "./output",
"specPath": "./specs/demo-review.md",
"focus": ["caption readability", "failure trace"]
}
focus defaults to an empty array. specPath is optional and should point to a
local spec or brief the reviewer should consider.
Outputs
The result contains a grounded prompt assembled from discovered artifacts,
recommended focus areas, and optional spec/focus inputs. It is designed to be
handed to Codex or Claude Code for the next review step. When
timeline.evidence.json exists, the prompt includes a short timeline preview.
When video.shots.json exists, the prompt includes a short shot-structure
preview. When segment.evidence.json exists, the prompt includes a short
segment-evidence preview. When canonical analyzer files exist at the bundle
root, such as analyzer.report.json, media-probe.json, quality-gates.json,
or caption-artifact.json, the prompt lists them as report evidence so the
reviewer starts from the shared analyzer output.
The prompt also routes common reviewer failure modes:
quality-gates.json failures and warnings, with first timestamped evidence
quality.json failures and warnings from consumer repos such as demo-machine
demo-visual-review.diff.json and golden-frame.diff.json mismatch summary
- legacy
visual-diff-report.json regression summaries when present
demo-capture-evidence.json or screenshots/manifest.json screenshot evidence
caption-artifact.json readability or sync risk plus caption-related gates
layout-safety.report.json caption, safe-zone, OCR, and overlap issues
- an explicit suggested first artifact to inspect next
Sequencing Guidance
Run after video-artifact-intake or review-bundle if you need to delegate a
review. If storyboard artifacts are needed but absent, run
storyboard-extract, storyboard-ocr, and optionally
storyboard-transitions/storyboard-understand before packaging the prompt.
Interpretation Notes
The generated prompt is a routing aid, not a review verdict. It should point the
next agent at real files and questions. If the bundle is sparse, the prompt
should preserve uncertainty rather than invent missing evidence.
Use timeline evidence to ground ordering, caption, transcript, and action claims
before relying on sparse storyboard frames.
Use shot evidence to choose which video segments or representative frames need
closer inspection. Do not treat shot boundaries as semantic proof by themselves.
Use segment evidence as the fastest routing layer for which parts have usable,
weak, or empty evidence.
Abstention Rules
Do not use this skill to summarize artifacts you have not generated or found.
Do not ask a reviewer to evaluate claims that are unsupported by the bundle,
unless they are explicitly framed as open questions.
Failure Modes
Wrong bundle locators produce a grounded prompt for the wrong run. Missing
specPath should not fail the run if omitted, but an invalid provided path can
make the prompt incomplete or misleading. If discovered artifacts are sparse,
run intake/storyboard steps first.
Repo-side runner:
node --import tsx scripts/harness/package-review-prompt.ts
Installed-pack runner:
node ./.video-evaluator/agent/run-tool.mjs package-review-prompt