The content-review Skill

---
name: content-review
description: Review supplied viewer comments, comparable performance figures and a timed script or shot plan, then combine the evidence into a useful local video revision brief.
---

# Review a video plan through three useful jobs

Use the supplied files to decide what the next version should explain or show more clearly. This method needs no social account or plugin. Preserve labels identifying fictional or prepared teaching data. A script and shot plan support a review of those documents, not a claim that you watched footage.

## Choose the assigned job

If you are an agent assigned the audience, performance or script review, perform only that role below. Use the inputs and output path in your assignment. Do not launch more agents, change another role's files or update the shared progress checklist. Report missing information to the main conversation rather than inventing an answer.

If you are in the main conversation directing the whole review, confirm the revision question and named input files. Use the existing project-local `progress.md` when supplied, or create a short checklist in the agreed review folder. The main conversation owns that file. Delegate three distinct roles when delegation is useful and available; otherwise perform the roles in sequence and say how the work ran. Delegation is part of this requested review, but creating reusable agent definitions or installing tooling is unnecessary.

Give each role its own inputs and output path. Keep the main conversation responsible for inspecting the returned files, updating progress and combining findings. Use these default filenames inside the agreed review folder unless the learner supplies others:

- Audience: `viewer-question-analysis.md`
- Performance: `performance-comparison.md`
- Script: `script-shot-review.md`
- Main synthesis: `revised-video-brief.md`

Read existing output files before replacing them. Resolve an unclear review folder with the learner before writing. Preserve the source comments, metrics and script.

An assigned agent should save its report at its assigned output path when its tools and permissions allow. If it cannot save, return the full report text and intended output path to the main conversation, explicitly stating that the file has not been saved and why. Do not claim a write succeeded or seek extra access merely to produce the report. Keep the same role boundaries and do not launch another agent to save it.

## Audience agent

Read the supplied comments, normally `viewer-comments.csv`. Identify the viewer questions and requests relevant to the revision goal. Keep comment IDs and short exact excerpts beside each finding. Group closely related comments and count distinct supporting IDs rather than repeated words. Distinguish an expressed question or request from your inference and from a compliment. Do not treat a small supplied sample as the whole audience.

Return the main questions, supporting IDs/excerpts, any useful disagreements and missing information in the audience output. Suggest what the next version could clarify, without claiming that the suggestion has been tested.

## Performance agent

Read the supplied metrics, normally `performance.csv`. Check observation age, video duration, units and missing values before comparing rows. Use comparable observation windows; explain any duration difference before comparing watch seconds. Do not silently mix unlike rows or treat absent data as zero.

Show the actual figures, formula and result for useful comparisons. Average watch as a share of duration is `average_watch_seconds / duration_seconds × 100`. When comparing the current item with several previous items, state which rows you used and whether the baseline is an arithmetic mean per video or a weighted measure. Do not weight watch time by views without explaining that choice. Distinguish a percentage-point difference in percentages from a relative percentage change. Keep precision reasonable and name each denominator.

Return a compact comparison table and the useful interpretation in the performance output. Views, watch time and the percentage remaining at a stated time answer different questions. A retention difference can identify a point worth investigating; these figures alone do not prove why viewers left or establish what a revision will achieve.

## Script agent

Read the supplied timed script and shot plan, normally `video-script.md`. Examine the order of information, named parts, clarity of actions, allotted time and whether the closing request serves the revision goal. Cite the actual time ranges and short script excerpts. Describe what the plan specifies, such as "the plan allocates the opening eight seconds to colour shots."

Suggest concrete changes to words, shots or timing while respecting the requested total duration. Report gaps or overlapping time ranges when present. Do not claim to have watched a shot, heard delivery or observed motion when only a plan was supplied. Save the assessment in the script output; leave cross-source priorities to the main conversation.

## Main synthesis and completion

When a role returns report text instead of a saved file, the main conversation saves that text unchanged at the intended local output path, reads the saved file back and checks that it is usable before marking the stage complete. Preserve who produced the findings and report that the main conversation saved them. If the report needs a substantive repair, request a focused repair from the assigned role rather than silently changing its findings. If the main conversation also cannot save, report the unsaved result and leave that stage unfinished.

Inspect the actual role outputs before checking their stages complete. Combine related evidence into the revision brief: what to explain or show, which comment IDs, metrics or script times support the change, and why the change could help. Separate observations from hypotheses. Reconcile disagreements, keep recommendations proportionate to the supplied evidence and leave final editorial choices to the learner. Do not edit a product page, produce a video or publish the revision as part of this review.

Only the main conversation updates `progress.md`. Mark a stage complete when the requested usable result exists. Record missing inputs and partial work with the path and next action. Each role reports what it completed, its output location and any unfinished item to the main conversation. The main conversation then tells the learner where to open `revised-video-brief.md`, what finished and what remains open. A successful process exit without the requested output is unfinished work.