Docs/Skills and code/skill video-script-production

.claude/skills/video-script-production/references/reader-prompts.md

Markdown52 lines5.9 KBSkill: video-script-production

What the file says about itself

Reader prompts (Step S10): Blind readers are the check on what the linter cannot see: whether a care line lands, whether a BUT really reverses an expectation, where a person would stop watching.

Mentioned on: page 06, page 12

Reader prompts (Step S10)

Blind readers are the check on what the linter cannot see: whether a care line lands, whether a BUT really reverses an expectation, where a person would stop watching. They need subagents, which only the user can authorise. In the Moser Baer run the user said "yes, run the fresh-reader pass on v4"; v5 and v6 have not had one.

How to run it

  1. Make the blind file: node scripts/extract-vo.js --script SCRIPT-vN.md --mode blind --out <scratch>/vo-only.txt. It holds numbered paragraphs only: no headings, no graphics, no chain, no lint output, no notes.
  2. Spawn two general-purpose subagents with model: "sonnet", in parallel: the viewer and the editor. Give each only the file path and the title and thumbnail description. Add a third reader only if the first two disagree on something important.
  3. Tell them to return the report as text in their final message (the harness rejects subagents that write report files). "Do not read ahead" is an honour-system instruction; say so in the hand-off.
  4. Read the reports, compare with the declared chain, and fix only what a reader backs with a quote. Each place where a reader disagrees with the chain (a BUT read as "and then", a care line that did not land) is a defect in the script, not in the reader. Mark any fix that touches a fact.
  5. Re-run check.js; log what the readers found, and what the linter had scored 100 but they still caught, in learnings.md.

Limits to state in the hand-off: two readers, one run, the same model family as the author, honour-system blindness.

Prompt 1: the viewer

You are a first-time YouTube viewer who knows nothing about <TOPIC> beyond <WHAT A TYPICAL VIEWER KNOWS>. You clicked this video: title "<TITLE>", thumbnail: <THUMBNAIL DESCRIPTION>.

Read ONLY this file, once, top to bottom, in order, as spoken narration (paragraphs are numbered): <PATH TO vo-only.txt>
Do not open any other file in that directory or anywhere else. Do not look anything up. React honestly as the viewer; do not be kind. Do not read ahead: write your notes for each paragraph before moving to the next.

Return your report as text, with:
1. Per paragraph (all <N>), one line: what question is in your head right now (or "none"), and your urge to keep watching 0-10. Flag any paragraph where you would plausibly click away, with why.
2. Confusion log: every term, number, name or step you did not follow on first hearing, with paragraph number.
3. Numbers: which figures you remember at the end, and which you heard but did not care about.
4. Joins: for each place the topic shifts (about <N JOINS>), say whether it felt like "therefore", "but", or "and then / next topic", quoting the words that made you feel it.
5. First 60 seconds (~150 words): by the end, what do you expect the video to give you, did it confirm the thumbnail, and do you feel it promises MORE than the title? Quote.
6. In your own words, the video's argument in two sentences, written from memory only. Then: the single weakest moment, and the single strongest.
7. Texture (added for v6): who is the narrator, in three adjectives, and would you trust them? Which comparison or story do you remember, and did it help or get in the way? Who were the people in the story, and which one would you describe to a friend? Did the first two minutes feel like a menu of things to come, or a story already under way?
8. Which sentences sound written rather than spoken, and the paragraph at which you stopped caring what happens next (if any).

Prompt 2: the skeptical editor

You are a sharp, skeptical editor reviewing a YouTube explainer voice-over for logic, flow and honesty. Read ONLY this file: <PATH TO vo-only.txt> (<N> numbered paragraphs, spoken text). Title: "<TITLE>". Do not open any other file anywhere; do not search the web.

Judge as an editor who has never seen any outline, and return your report as text:
1. Argument: state the thesis in one sentence as the script actually supports it. List each causal claim ("X led to Y") and rate it supported-in-the-script / asserted-only / leaps. Does the script's evidence justify the strongest word in the title?
2. Flow: list every section boundary. For each, apply the swap test (would swapping this section with the next break the logic?) and the "and then" test (could you insert "and then" between them with nothing lost?). Quote the sentences on each side of the weakest 3 joins and propose a fix.
3. Load: find any 30-second stretch (~75 words) that crams in too many figures, names or new concepts. Quote it. Find any figure that appears before the viewer has a reason to care.
4. Promises: list every promise or tease made in the first ~150 words and where (paragraph) it is paid off or dropped. Anything dropped or paid weakly?
5. Repetition, padding or any place a viewer already knows what is coming; any sentence that is vague or cliche.
6. Claims about people: any statement about what someone thought, wanted or intended, or any allegation, that goes beyond what the script itself shows.
7. Verdict: top 5 concrete edits, ranked, each with the paragraph number and a rewritten line.

What the v4 readers found (so you know what to expect)

The linter scored v4 100/100. The viewer recalled the argument correctly and rated the hook and the "well-funded" press conference highest; the editor judged that the evidence supported "the financing made the decline fatal", not "the escape plan killed it" (which backed the title change). Both caught what the linter could not: two paragraphs too dense in names and terms (the CARE rating paragraph and the CBI paragraph), an oversold tease, a link built on an adjacency the script said it could not support, a scam comparison that showed scale and not innocence, and a close that restated the thesis a third time. They also found jargon with no gloss (Philips, gigawatt, tribunal, Alchemist), which led to the glosses in v5 (rule A7 already checked declared jargon, and A8 flags undeclared watch-list terms).

Show the raw text
# Reader prompts (Step S10)

Blind readers are the check on what the linter cannot see: whether a care line lands, whether a BUT really reverses an expectation, where a person would stop watching. They need subagents, which only the user can authorise. In the Moser Baer run the user said "yes, run the fresh-reader pass on v4"; v5 and v6 have not had one.

## How to run it

1. Make the blind file: `node scripts/extract-vo.js --script SCRIPT-vN.md --mode blind --out <scratch>/vo-only.txt`. It holds numbered paragraphs only: no headings, no graphics, no chain, no lint output, no notes.
2. Spawn two `general-purpose` subagents with `model: "sonnet"`, in parallel: the **viewer** and the **editor**. Give each only the file path and the title and thumbnail description. Add a third reader only if the first two disagree on something important.
3. Tell them to **return the report as text** in their final message (the harness rejects subagents that write report files). "Do not read ahead" is an honour-system instruction; say so in the hand-off.
4. Read the reports, compare with the declared chain, and fix **only what a reader backs with a quote**. Each place where a reader disagrees with the chain (a BUT read as "and then", a care line that did not land) is a defect in the script, not in the reader. Mark any fix that touches a fact.
5. Re-run `check.js`; log what the readers found, and what the linter had scored 100 but they still caught, in `learnings.md`.

Limits to state in the hand-off: two readers, one run, the same model family as the author, honour-system blindness.

## Prompt 1: the viewer

```
You are a first-time YouTube viewer who knows nothing about <TOPIC> beyond <WHAT A TYPICAL VIEWER KNOWS>. You clicked this video: title "<TITLE>", thumbnail: <THUMBNAIL DESCRIPTION>.

Read ONLY this file, once, top to bottom, in order, as spoken narration (paragraphs are numbered): <PATH TO vo-only.txt>
Do not open any other file in that directory or anywhere else. Do not look anything up. React honestly as the viewer; do not be kind. Do not read ahead: write your notes for each paragraph before moving to the next.

Return your report as text, with:
1. Per paragraph (all <N>), one line: what question is in your head right now (or "none"), and your urge to keep watching 0-10. Flag any paragraph where you would plausibly click away, with why.
2. Confusion log: every term, number, name or step you did not follow on first hearing, with paragraph number.
3. Numbers: which figures you remember at the end, and which you heard but did not care about.
4. Joins: for each place the topic shifts (about <N JOINS>), say whether it felt like "therefore", "but", or "and then / next topic", quoting the words that made you feel it.
5. First 60 seconds (~150 words): by the end, what do you expect the video to give you, did it confirm the thumbnail, and do you feel it promises MORE than the title? Quote.
6. In your own words, the video's argument in two sentences, written from memory only. Then: the single weakest moment, and the single strongest.
7. Texture (added for v6): who is the narrator, in three adjectives, and would you trust them? Which comparison or story do you remember, and did it help or get in the way? Who were the people in the story, and which one would you describe to a friend? Did the first two minutes feel like a menu of things to come, or a story already under way?
8. Which sentences sound written rather than spoken, and the paragraph at which you stopped caring what happens next (if any).
```

## Prompt 2: the skeptical editor

```
You are a sharp, skeptical editor reviewing a YouTube explainer voice-over for logic, flow and honesty. Read ONLY this file: <PATH TO vo-only.txt> (<N> numbered paragraphs, spoken text). Title: "<TITLE>". Do not open any other file anywhere; do not search the web.

Judge as an editor who has never seen any outline, and return your report as text:
1. Argument: state the thesis in one sentence as the script actually supports it. List each causal claim ("X led to Y") and rate it supported-in-the-script / asserted-only / leaps. Does the script's evidence justify the strongest word in the title?
2. Flow: list every section boundary. For each, apply the swap test (would swapping this section with the next break the logic?) and the "and then" test (could you insert "and then" between them with nothing lost?). Quote the sentences on each side of the weakest 3 joins and propose a fix.
3. Load: find any 30-second stretch (~75 words) that crams in too many figures, names or new concepts. Quote it. Find any figure that appears before the viewer has a reason to care.
4. Promises: list every promise or tease made in the first ~150 words and where (paragraph) it is paid off or dropped. Anything dropped or paid weakly?
5. Repetition, padding or any place a viewer already knows what is coming; any sentence that is vague or cliche.
6. Claims about people: any statement about what someone thought, wanted or intended, or any allegation, that goes beyond what the script itself shows.
7. Verdict: top 5 concrete edits, ranked, each with the paragraph number and a rewritten line.
```

## What the v4 readers found (so you know what to expect)

The linter scored v4 100/100. The viewer recalled the argument correctly and rated the hook and the "well-funded" press conference highest; the editor judged that the evidence supported "the financing made the decline fatal", not "the escape plan killed it" (which backed the title change). Both caught what the linter could not: two paragraphs too dense in names and terms (the CARE rating paragraph and the CBI paragraph), an oversold tease, a link built on an adjacency the script said it could not support, a scam comparison that showed scale and not innocence, and a close that restated the thesis a third time. They also found jargon with no gloss (Philips, gigawatt, tribunal, Alchemist), which led to the glosses in v5 (rule A7 already checked declared jargon, and A8 flags undeclared watch-list terms).

Shown as it is in the repository at build time (2026-11-03); the home directory appears as <home>. If this page and the file differ, the file is right. Back to the catalogue.