How to go from a raw phone recording to a finished Reel or TikTok, with captions, motion graphics, a podcast-quality voice and three thumbnails. What you need, the prompts to paste, and how to check the result.
Same kit, same process, three different looks. Each one started as a plain selfie video. The agent wrote the plan, cut the pauses and retakes, built the motion graphics, processed the voice and designed the thumbnails.
80 s · teal brand stage, presenter card, logo that builds itself, price strike-through.
60 s · alert red and green, split screen, a real headline as proof, code that lights up red.
28 s · glass and lime, every effect he names happens on that word, 3D layers, comment CTA.
Five things. The first run on a new computer installs the tools (about 15 minutes); after that every video starts straight away.
The skills in project-template/.agents/skills (or the Video Montage plugin once it is released). The director skill video-director is the entry point for every video.
For the 3 thumbnails (Nano Banana Pro, about 75 credits each) and any AI image or clip. Turn the connector on in Agent Cowork before you start the job.
Vertical phone video, 1080 × 1920, 30 fps. One or more clips. Photos, logos and links are welcome too.
Node 22+, Python 3.10+, ffmpeg with rubberband, 10 GB free. The agent installs Remotion, Whisper models and OpenCV 4.10 on the first run.
What should the viewer do at the end? "Comment PROMPT", "book a visit", "follow for part 2". The whole edit is built around it.
A brand file (colours, fonts, logo) in brands/. With it, the prompt can simply say "use the B Premier style" and every video stays on brand.
| Per video | Typical |
|---|---|
| Time, unattended | 60 to 90 minutes for a 30 to 90 s Reel |
| Magnific credits | about 225 for 3 thumbnails, more only if you ask for AI images or clips |
| Your time | 5 minutes to set up, 5 minutes to review |
The agent can remove pauses, retakes and "uhh", but it cannot fix a dark room or a missing sentence. Eight habits that make the biggest difference:
Light in front of you. A window or lamp facing you, never behind you.
Quiet room, phone close. Turn off fans and AC if you can. Record 2 seconds of silence first.
Phone at eye level, vertical, with space above your head for titles.
Stumbled? Pause one second and say the whole sentence again. The agent keeps the last complete take.
Name what you want to see. The director shows what you say: say "the price drops to 79 dollars" and the price will drop on screen.
Hook in the first sentence. Start with the most surprising line, not "Hi guys".
Say the call to action clearly, with the exact keyword: "comment PROMPT".
Wear a colour that contrasts with your background. A plain white top on a white wall makes everything harder.
"The zoom in", "the glass effect", "stop", "more like Codex": each frame shows the moment the words were said.
Copy project-template from the kit, rename it, drop your files in input, open it in Agent Cowork. That is the whole setup.
my-video-01/ ├─ AGENTS.md the agent reads this first ├─ CLAUDE.md same file, other agents ├─ START-PROMPT.txt the prompt to paste ├─ AGENT-RULES.md ├─ VIDEO-QUALITY-CONTRACT.md ├─ style-log.json looks already used ├─ input/ your clips, photos, logo ├─ brands/ optional style files ├─ deliveries/ the results land here └─ .agents/skills/ video-director + 11
Copy project-template and give it a name, for example reel-2026-10-07.
Put your files in input/. Name clips in the order they play: 01-main.mov, 02-cta.mov.
Optional: copy brands/_template.json to brands/your-brand.json and fill in colours and fonts.
Open the folder in Agent Cowork in a new chat. Check that Magnific shows as connected.
Paste the prompt from the next page, note the start time, and let it run.
This is the standard prompt for any video. It authorizes the full job, sets the quality bar and asks for a short report at the end.
Open this folder and follow AGENTS.md exactly. Make ONE finished vertical Reel/TikTok from the files in input/ with the video-director skill. The full job is authorized: work to the end without asking me. - Goal: [what the viewer should do, e.g. comment PROMPT] - Style: [brand file or colours, or "pick a new look"] - Language: [English / Arabic], Western digits 0-9 only - Voice: thick podcast voice - Every visible thing I mention must HAPPEN on screen at that word. A label or a list does not count. - Remove every pause, filler, false start and repeated phrase. Keep only one complete take of each sentence. - 3 thumbnails with Nano Banana Pro on Magnific: tell me the credit cost, then continue without waiting. - Do not deliver until qa_check.py prints PASS and the independent rubric is at least 40/50. When done, reply with: total time, qa_check result, rubric score, and the list of files in deliveries/.
Paste the standard prompt, then add the block that fits. All of these are in PROMPTS.txt in the kit.
Clips play in file-name order. If a sentence is started in one clip and said again in the next, keep only the complete take.
Style for this video: - Colours: bg #0A0A0A, main #FFD400, text #FFFFFF, warning #FF3B3B - Fonts: Anton (titles), Inter 800 - Mood: bold, punchy, TikTok energy Keep the readability and timing rules.
Language: Arabic. Captions right-to- left, animate whole words, Arabic font (Cairo or Tajawal) one weight heavier than English. Western digits 0-9 only. Transcribe with the language set to Arabic and check timings twice.
Type: real-estate promo from the photos in input/. Strongest view first, then location map, room by room, specs from input/specs.pdf only, price, contact. No generated rooms. Music, no voice.
Type: podcast-clip. Find the best 45 to 75 s moment with one clear idea and a hook line in the first 3 s. Show who is speaking. Deliver 3 options as clips.
Review deliveries/final.mp4 and fix only these, then re-run qa_check: 1. [at 0:04] the hook text is too small 2. [at 0:12] the glass panel covers my eyes 3. end card shorter (2 s)
The agent thinks like a director before it writes any code. Everything goes into one plan file, work/video-plan.json, and every step has a check before the next one starts.
Understand. Reads every file, transcribes the speech twice, writes a 6-line brief with one goal.
Classify. Talking head, podcast, promo, real estate, explainer, AI story… each type has its own rules.
Concept and look. One idea that makes this video special, plus colours, fonts and captions.
Beat plan. Second by second: what happens on screen. Every named thing becomes a cue.
Assets. Your files, your brand, official logos and screenshots, licensed icons and fonts, real sources.
Generate. Only what is missing, through Magnific, with the credit cost shown first.
Compose. Builds the video in Remotion: captions, motion graphics, layouts, transitions.
Sound. Thick podcast voice, one sound effect per visual event, loudness at −14 LUFS.
Quality check. Automatic checks plus a separate model that looks at the frames and scores them.
Deliver. Final video, 3 thumbnails, the plan, short notes with scores and credits.
| File | What it is |
|---|---|
| final.mp4 | The video. 1080 × 1920, H.264, ready to upload to Reels, TikTok and Shorts. |
| thumb-1/2/3.jpg | Three cover options. The title is set in code, so the text is always sharp and spelled right. |
| NOTES.md | One page: the look, the rubric scores, credits spent, what was simplified. |
| qa_check.txt | The automatic check. The first line must say PASS. |
| cues.csv + cues/ | Every thing you named, its time, and a frame proving it is on screen. |
| scene-map.csv | One row per visual event, useful when you ask for changes. |
| video-plan.json, RUN-LOG.md | The full plan and the real time spent on each step. |
PASS - runtime reduction 14.9% on speech length pauses were cut - first visual event at 0.00s (<=1.0) strong hook - fillers um/uh/mm (max 2): none clean speech - voiced gaps >= 0.8 s: none no hidden "uhhh" - caption words not heard: none captions = your words - numbers on screen never said: none nothing invented - picture after the last word 2.20 s end card - loudness -14.0 LUFS platform volume
The agent sends a sharp frame of your face to Nano Banana Pro as a reference, generates three different scenes with empty space at the top, then adds the title in code. You pick the one you like.
Real covers from the effects showcase (first three) and the security warning video.
Generated images misspell words. Titles are drawn afterwards, 2 to 5 words, large and bold.
Instagram's grid shows only the middle 3:4. Titles stay inside it, nothing important at the bottom.
About 75 credits per image. The agent states the total before generating.
By default the agent picks a style recipe it has not used in your last three videos. If you give colours, fonts or a brand file, those win.
| Recipe | Best for | Signature |
|---|---|---|
| brand-stage | Product reviews, launches | Logo builds itself, presenter card under a stage |
| threat-scan | Warnings, myths, security | Red/green split, code that lights up |
| glass-lab | AI tools, effects, tech demos | Frosted glass, 3D layers, effects counter |
| editorial-paper | Opinion, stories | Paper frames, marker underline |
| neon-terminal | Developer tools, CLI | Everything types in |
| bold-pop | Tips, "3 things" lists | Numbered stickers slam in |
| data-news | Numbers, studies, markets | Real article clips, counters |
| minimal-mono | Calm, premium, personal | One huge word at a time |
brands/my-brand.json
{
"name": "My Brand",
"palette": {
"bg": "#0A0A0A",
"primary": "#FFD400",
"text": "#FFFFFF"
},
"fonts": {
"display": "Anton",
"body": "Inter 800"
},
"logo": "brands/my-brand/logo.svg",
"tone": "confident, short"
}
Use the fix round prompt (page 7) with the time and the problem. Small fixes take 10 to 15 minutes, mostly the new render.
| You see | Do this |
|---|---|
| "Magnific is not connected" | Turn the connector on, start a new chat in the same folder, run again. |
| qa_check says FAIL, runtime below 0 % | The check was given one clip only. All clips must be joined for --original. |
| Labels instead of real effects | Make sure the prompt has the "HAPPEN on screen" line and that cues.csv exists. |
| Stripes or glitches in the render (Windows) | Ask the agent to switch from OffthreadVideo to Video and re-render. |
| The job runs over 2 hours | Open RUN-LOG.md to see the slow step; usually first-time installs or a render retry. |
| Every video looks the same | Keep style-log.json between videos, or give a style block. |
Record vertical, light in front, hook first, name what you want to see, clear call to action.
Copy project-template, put clips in input/ (in play order), add a brand file if you have one.
Open the folder in Agent Cowork, check Magnific is connected.
Paste the standard prompt, fill goal, style and language.
Wait 60 to 90 minutes. The director plans, sources, generates, builds and checks.
Review on your phone with the sound off; qa_check must say PASS.
Fix round if needed, then pick a thumbnail and post.