b-roll-planner

---
name: b-roll-planner
description: Turns a finished short-form script into a shot-by-shot filming plan and a batched shoot order. Use once a script exists and filming has not started.
---

# Know exactly what to film before you pick up the phone

A shot table matched to your script line by line, and a shoot order grouped to film one location, one outfit, one go.

## Claude does

1. Asks for four inputs, refusing to start without the first: the script or voiceover; target seconds; your available locations, props and kit; whether you are on camera, and how many outfits.
2. Splits it into beats on the pacing the sources agree: about 75 spoken words per 30 seconds, a visual change every 2 to 4 seconds. So 45 seconds means roughly 110 words and 12 to 20 shots.
3. Writes a two-column AV script, audio left, visuals right, tagged [VO], [SFX] and [Action].
4. Fixes the hook shot first. The opening 1 to 3 seconds must be filmable with what you own, and must show the script's promise, not describe it.
5. Marks every beat muted-first. Around 80 to 85% watch with sound off, so if the visuals alone do not tell it, Claude rewrites the shot, not the caption.
6. Refuses to guess. No location, prop, person or permission outside your list enters the plan; missing items are listed as "get this before filming".

## Then Codex does

Codex does the bulk arithmetic and reordering: every beat expanded into a row (shot number, seconds, framing, camera move, subject, prop, location, outfit), timings summed against the target, the table re-sorted by location, then framing, then outfit, plus a call sheet and kit list.

Claude fills in first: {AV_SCRIPT}, {TARGET_SECONDS}, {CONSTRAINTS} (locations, props, kit, outfits), {ON_CAMERA}.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. Every audio line has a visual beside it, and no shot holds past 4 seconds without reason.
2. Shot count and total seconds hit the target within 2 seconds.
3. Reads the visuals column alone. If the story does not land silently, it fails.
4. No row uses anything absent from {CONSTRAINTS}. Invented props are the commonest failure; a plan you cannot film is worse than none.
5. The shoot order never returns to a location already left.

Any check fails: hand those rows back to Codex with the reason and rerun, never patch by hand.

## Rules

- Frame vertical 1080x1920, subject clear of the burned-in captions. Safe zones differ per platform, so check a live phone preview.
- Benchmark numbers in skill repos are house numbers unless a study is named. Say which.
- Walk a generated shot table yourself before shooting it.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude builds both itself, and says so.

## Built from

- social-media-skills/skills, https://github.com/social-media-skills/skills, 66 stars, api.github.com, 5 Sep 2026. AV script, batching, pacing.
- AgriciDaniel/claude-shorts, https://github.com/AgriciDaniel/claude-shorts, 204 stars, api.github.com, 5 Sep 2026. The 1080x1920 specs.
- Jenny Hoyos, https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook, 10m views per Short, on YouTube's own blog. Hook under 3s.
- MrBeast handbook, https://simonwillison.net/2024/Sep/15/how-to-succeed-in-mrbeast-production/, leaked internal document. Every second justifies itself.

# Know exactly what to film before you pick up the phone - Codex task

## You are given

- {AV_SCRIPT}: a two-column audio/visual script already agreed with the user. Audio lines are tagged [VO] or [SFX]; visual lines are tagged [Action]. Every audio line already has a visual beside it.
- {TARGET_SECONDS}: the finished video length, typically 30 to 60.
- {CONSTRAINTS}: the only locations, props, kit and outfits available. Nothing outside this list may appear anywhere in your output.
- {ON_CAMERA}: yes or no, and how many outfits.

## Produce

Three files in the working folder.

1. `shot-table.csv`, one row per shot, columns exactly: `shot_no,seconds,framing,camera_move,subject,prop,location,outfit,audio_line,visual_action`.
   - Shot count: {TARGET_SECONDS} divided by 3, rounded, plus or minus 30%. For 45 seconds that is 12 to 20 rows.
   - `seconds` is 2 to 4 per shot unless the script marks a deliberate hold. The column must sum to {TARGET_SECONDS} plus or minus 2.
   - `framing` is one of: wide, mid, close, extreme close, over-shoulder, top-down, screen recording.
   - `camera_move` is one of: static, pan, tilt, push in, pull out, handheld follow.
   - Shot 1 is the hook and runs no longer than 3 seconds.
2. `shoot-order.md`, the same rows re-sorted for filming, grouped by location first, then framing, then outfit. Each group has a heading and the shot numbers it covers. A location must appear once only. Note under each group what has to be reset or moved before the next group.
3. `call-sheet.md`, one page: shoot date if given, the group order with a running time estimate, a kit list built only from {CONSTRAINTS}, and a "missing before filming" list of anything the script needs that {CONSTRAINTS} does not contain.

## Rules

- Never add a location, prop, person or outfit that is not in {CONSTRAINTS}. If a visual line needs one, keep the row, leave the field as `MISSING`, and add it to the missing list.
- Never rewrite the audio lines. You reorder and expand visuals only.
- Frame everything vertical 1080x1920. Keep the subject clear of the lower third where burned-in captions sit.
- Each visual must read with the sound off. If a row only works with narration, flag it in a `WEAK-MUTED` column entry rather than inventing a better shot.
- No em dash characters. British English.
- Do not invent a fact, a number or a quote.

## Return

Print to stdout, in this order:

1. The full `shot-table.csv` contents.
2. `TOTAL SECONDS: <sum> vs TARGET: {TARGET_SECONDS}`.
3. `SHOT COUNT: <n>`.
4. `LOCATION GROUPS: <n>` and the group order on one line.
5. Every `MISSING` and `WEAK-MUTED` row number, or `NONE`.
6. The three file paths written.