Communitygithub.com

MonsterDylan/demo-reel-plan

Plan a short-form demo reel for an AI system, tool, workflow or build - a beat sheet plus a shoot and edit spec - using techniques ranked by real performance across 66 saved reels from 60 creators. Use this whenever someone wants to plan, outline, storyboard, structure or script a Reel, TikTok or Short that shows something they built or a tool in action (a screen tour, a tutorial, a "watch my AI do X" demo), or asks how to film or edit one, even if they only say "help me make a reel about my system".

What is demo-reel-plan?

demo-reel-plan is a Claude Code agent skill that plan a short-form demo reel for an AI system, tool, workflow or build - a beat sheet plus a shoot and edit spec - using techniques ranked by real performance across 66 saved reels from 60 creators. Use this whenever someone wants to plan, outline, storyboard, structure or script a Reel, TikTok or Short that shows something they built or a tool in action (a screen tour, a tutorial, a "watch my AI do X" demo), or asks how to film or edit one, even if they only say "help me make a reel about my system".

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/MonsterDylan/claude-saved-to-skills/tree/HEAD/examples/ai-content-ideas/skills/demo-reel-plan

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Demo reel plan

This plans a reel that shows an AI system or tool doing something. The output is a beat sheet and a shoot and edit spec the person can film from.

The main finding behind it: in the collection this is built from, the reels that travelled furthest made the demo feel live. A visible person, a real screen filmed by hand, something going slightly wrong and being handled. The heavily produced techniques (a graphic on every word, designed mock screens) sat below the median. So this skill pushes toward showing the real thing happening, and treats polish as optional.

Evidence for every rule is in references/evidence.md. Examples are in references/examples.md. Lift is points above or below the median of comparable reels on a 0 to 100 score. It is correlation in one hand-picked collection of 66, so it ranks options. It does not guarantee anything.

The rules that showed a lift

Start mid-action. +20, 7 creators. Open on the task starting, not on an introduction. The first words are the command, the prompt, or the first step. [DboqOSvEj_b]

Keep a face in frame during the screen demo. +16, 8 creators. A small bubble, a lower panel, a reaction layer. The screen never fully replaces the person. A demo with no human in it reads as an ad. [Dd6izTYiln1, DboqOSvEj_b, DeHX34IxveJ]

Put friction inside the demo. +12, 6 creators. One moment where it does not go smoothly: the slot is taken, someone asks if it is an AI, the task fails without the tool and then works with it. Friction gives a demo a middle. Plan it, and put it around the 40 percent mark. Use a real one from the build if there is one. [DboqOSvEj_b, DdrUdGbkrT0, Dd6izTYiln1]

Film the real screen with a phone. +8, 11 creators. Handheld, pointing at it, pushing in on the result. It proves the thing exists and is running. Use a screen recording only for text the viewer must read. [DbcPo0dMQ3d, DbYhDu_MADd, DXA4gi1DCOx]

Show the technique on the video itself. +18, 3 creators. Thin evidence, use when it applies. If the reel teaches an edit, a prompt or a structure, the reel should visibly use it. [De8SPjEAuV0, DambEp1igDW]

Show the source at the moment of the claim. +13, 3 creators. Thin evidence. The repo page, the headline, the README, on screen when you say it. [DfCXigRSyAo, Dd6izTYiln1]

Defaults (everyone does these, no lift either way)

  • First frame shows the finished thing working (19 creators).
  • When face and screen share the frame: screen or graphic on top, face below, captions on the seam (17 creators).
  • Word-by-word captions, one to three words, one accent colour (16 creators).
  • Say the step count and label each step on screen (18 creators).
  • A marked late beat, "here's the trick" or "the crazy part", before the ask (8 creators, +4).
  • Prove it is real early: click into it, open the file (9 creators, +4).
  • One comment keyword as the last spoken line, with the keyword on screen (32 creators). It raises comments (2.73 percent versus 0.16 percent) and does not change reach.

Cautions (these sat below the median)

  • A new visual on every spoken noun, every 1 to 3 seconds. -23, 15 creators. The most common heavy edit and the weakest result. Use graphics where they carry information the footage cannot. Do not carpet the reel.
  • Designed mock screens in place of the real product. -25, 5 creators. Show the real thing, even if it is uglier.
  • Cutting back to your face for the ask. -25, 8 creators. Keep the product on screen and pin the keyword as a label. If the face is already in frame, stay in that shot.
  • Touring a system one area at a time. -16, 6 creators. A tour is a list. The one that worked opened on a counted claim, proved the agents were clickable in 10 seconds, and added a second reason to care mid-reel. If the plan is a tour, give it a task to follow instead: one job moving through the system.
  • Showing the deliverable right before the ask. -16, 6 creators. If you show the freebie, show it earlier, at the point the viewer would want it.

Length

Do not cut for shortness alone. Teach reels in the top third had a median length of 63 seconds, the bottom third 42. Reels over 60 seconds scored higher than reels of 45 or under (56.5 versus 43.5). Let the demo run as long as it stays live, and cut dead air, not substance. Hard-cut count showed no pattern either way.

What to produce

# <working title>
Promise: one sentence, what the viewer sees happen
Length:  target in seconds, with the reason
Proof:   what on screen shows this is real

## Beats
| Time | Beat | Says | On screen | Shot |
(open mid-action, prove it is real by 0:10, friction near 40 percent, payoff, late beat, ask)

## Shoot list
What to film, how (phone on screen, face bubble, screen recording), in the order it is easiest to capture

## Edit spec
Layout, captions, the few graphics that earn a place and why, what is deliberately left raw

## Ask
The last spoken line and the on-screen label

## Before filming
What must not be on screen: private data, client or customer names, tokens, keys, file paths

## Evidence used
Only the rules that changed a decision in this plan, one line each, 8 lines at most

A demo of a real system shows real data. Check what the screens will reveal (names in a sidebar, a token in a terminal, the content of an answer) and say how to keep it out of frame. This is easy to miss and expensive to get wrong.

Keep the evidence section short. The person needs to know why the plan differs from what they would have done, not a restatement of this file.

If the person also wants a caption, a giveaway outline or a delivery message, those are separate jobs. Say so in one line and offer them, so the plan stays a plan.

Write the beats so someone could film them today. Use the person's real system and real numbers. If you do not know what something looks like on their screen, ask or mark the assumption. If no friction moment exists, propose one honest candidate and say it needs to be real.

Fit with the user's own rules and data

If the user has their own brand rules, edit standard or performance data from their own account, those win over this collection, because their audience is not the same as these 60 creators. Where a rule here disagrees with theirs, say so in one line and let them choose. Do not silently apply either.

Related Skills