My Daily AI Stack.

How I use Claude, ChatGPT, Gemini and Copilot, by task. Every step shows the tool I use for it.

1. Reels

Claude

HeyGen

Codex

CapCut

Also

Claude Design

Gemini Veo

ChatGPT

From script to posted reel, with the comments handled after. Codex does most of the video work.

1

Step

Start from the idea

Claude

How I start depends on the type of reel.
AI tutorial videos: I test the tool over and over until I know it inside out. Then I brainstorm the angle with Claude.
Real-life takes: I start with my own idea. I save references to WhatsApp as I find them, then work out with Claude how each part fits in.
AI updates: Speed matters most, so I go straight from the update to a script with Claude (step 2).
Either way, I pick the take most people don't have, not a summary of the news.

Brainstorm angles

I want to post about [topic / tool / feature].
Audience: [who it's for].
What I have so far: [what I found testing it, or my saved
references].
1. Give me 10 angles, from safest to boldest.
2. For each one: a one-line hook and why it would stop someone
   scrolling.
3. Mark which work best as a reel, a carousel or a post.
4. Skip anything generic. If an angle depends on a fact you're
   not sure of, say so.
Then ask me which one to develop.

2

Step

Write the script

Claude

I've created my own script skill, based on how I write reels, called reel-script-writer. Create yours with the "Create your own skill" prompt at the end of this guide.
Then I type /reel-script-writer with the angle. Hook first, real steps, one clear payoff.
It ends with a keyword to comment, like "guide", and Claude writes the caption too.
No skill yet? The prompt below does the same job.

Reel script

Write an Instagram reel script about [topic / tool / feature].
Rules:
1. First word is the name of the tool the reel is about.
2. Hook in the first line: the single most surprising thing.
3. If it's a tutorial, give the exact steps in order. Real menu
   names only. If you don't know a step, ask me, don't guess.
4. One big reveal moment where the payoff lands.
5. Payoff in short, specific fragments. Real numbers, no adjectives.
6. 150 to 220 words. One continuous script, no labels, no emojis,
   no stage directions.
7. No em dashes. Never use: leverage, unlock, seamless,
   game-changer, delve, elevate.
8. End with: Comment 'Guide' and I will send you the exact steps
   and the prompt to use.

3

Step

Record it with my AI avatar

HeyGen

I made my avatar, a digital twin of me, once in HeyGen by recording a short training video of myself.
Now I paste each script into HeyGen and the avatar reads it.
That's why the script comes out as one clean block with no labels.
For some reels I record myself on camera instead.

4

Step

Clean up and edit

Codex

CapCut

Codex cuts the pauses and dead air, then speeds it up to 1.2x.
I've saved this as a skill in Codex. The prompt below does the same job, and you can save it as your own skill.
For a new video, I first ask Codex to watch it end to end and suggest how to edit it.
Once the b-roll is ready (next step), I do the final edit in CapCut.

Clean up the recording

Clean up [video file].
1. Remove silent pauses and dead air. Keep a tiny gap at each cut
   so words don't sound chopped.
2. Speed it up to 1.2x.
3. Save it as a new file. Never overwrite the original.
4. Check the first 10 seconds and two later cuts, then tell me
   the final length and size.

5

Step

Make the b-roll

Codex

Claude Design

Gemini Veo

Codex makes end-to-end b-roll matched to the script, using HyperFrames and Remotion (free tools that build videos from code).
Install them once by asking Codex: "Install the HyperFrames and Remotion skills."
I keep reusable logos and branded backgrounds in one folder so every reel looks consistent.
Claude Design: animated scenes of app screens, steps and numbers.
Gemini Veo: cinematic, real-looking clips.

End-to-end b-roll

Create end-to-end b-roll for [video file].
1. First watch the whole video and plan it against what I say.
   Show me the plan before building anything.
2. Match every visual to the script, not random footage.
3. Change the visual every 3 to 5 seconds.
4. Make two versions: full-screen b-roll, and 35% screen b-roll
   I can place over my talking head.
5. Keep all text and elements inside the Instagram safe zone.
6. Use my brand assets from [folder]: logos, backgrounds, colours.
7. No captions, no music. Light sound effects are fine.
8. Don't edit the source footage. Check that voice and visuals
   stay in sync to the end.

Claude Design, b-roll from a script

Here is my reel script: [paste script].
Create animated b-roll for it.
1. Size: 1080 x 1920, 9:16 vertical.
2. Split the script into scenes, one scene per sentence or beat.
   Time each scene to match speaking pace, about 2.5 words per
   second.
3. Show what the script talks about: the real app screens, the
   steps, the numbers. Animate key numbers and phrases as bold
   on-screen text.
4. Style: [background colour], [accent colour], [font style].
5. No voiceover, no people, no music.
6. Export as MP4.

Veo b-roll

A [length]-second vertical 9:16 video.
Subject: [what's in the shot].
Setting: [where].
Camera: [slow push in / slow pan left / static].
Lighting: [golden hour / night, screen glow / soft daylight].
Mood: [calm / focused / premium].
No text, no logos, no recognisable faces.

6

Step

Make the thumbnail

Claude

ChatGPT

I've saved my thumbnail style as a skill, so Claude writes the thumbnail prompt from the script. The prompt below works without a skill.
I attach my photo in ChatGPT and paste the prompt.
Same layout every time: tool logo at the top, punchline in a red box at the bottom.

Reel thumbnail

Create a reel cover. PORTRAIT, 1080 pixels wide by 1920 pixels
tall, 9:16.
Use the attached photo of me. Keep my face exactly as it is.
Tight chest-up crop, looking at the camera, [expression].
BACKGROUND: dark with a subtle glowing circuit pattern.
TOP: the official [tool] logo, large and accurate.
Two small dark rounded icon badges beside me: [icon 1], [icon 2].
BOTTOM: a bold red box with white text: "[punchline, max 5 words]"
AVOID: changing my face, fake logos, extra text, misspelled text.

7

Step

Post, upload and reply to comments

Instagram

Codex

YouTube

Facebook

I post on Instagram first, by hand, and cross-post it to Facebook from Instagram.
Then Codex uploads it to YouTube as a Short with a how-to title, description and tags. It works through my logged-in YouTube Studio in Chrome, using an upload skill I saved.
Codex also replies to "Guide" comments on Facebook and YouTube with the right guide link. I built that reply skill with the prompt below.
Pricing questions, business enquiries and complaints are flagged for me, not auto-replied to.

Comment reply skill for an agent

Create a skill for replying to comments on my [platform] posts.
1. My voice: [3 words, e.g. friendly, direct, a little dry].
   Here are 5 replies I've written that sound like me:
   [paste examples].
2. Keep replies short: 1 or 2 sentences.
3. If someone comments "Guide", reply with: [your guide reply
   and link].
4. Answer real questions directly. If you don't know the answer,
   say I'll get back to them.
5. Never argue. Skip spam and bots.
6. Don't reply to anything about pricing, business enquiries or
   complaints. Flag those to me instead.

↑ Back to contents

2. Carousels

Claude

ChatGPT

Claude writes the words and the image prompts. ChatGPT makes the slides.

1

Step

Write the slide-by-slide copy

Claude

I start with the same brainstorm prompt as reels and pick an angle.
I've saved my carousel rules as my own skill, called carousel-builder. Create yours with the "Create your own skill" prompt at the end of this guide. Or use the prompt below as it is.
Then I type /carousel-builder with the angle I picked.
It gives me the slide copy, 10 cover lines and the caption.
The last slide asks people to comment a keyword, like "Prompt", to get the prompt.

Carousel copy

Build an Instagram carousel on [topic].
My angle: [the take most people don't have].
Audience: [who it's for].
Rules:
1. Cover: max 8 words. A promise or a number, never a topic label.
2. Slide 2 proves it works. Show the result first, not the problem.
3. One idea per slide. Every slide must make sense as a screenshot
   on its own.
4. One save-worthy slide (a prompt, checklist or framework). Flag it.
5. Headlines state the lesson directly. No build-up.
6. Match the real step count. Never pad to a rounder number.
Give me: slide-by-slide copy, 10 cover lines from safest to
boldest, and a caption that starts with the CTA and adds one thing
the slides don't say. Max 5 hashtags.

2

Step

Turn each slide into a ChatGPT prompt

Claude

I've saved this as a skill too. The prompt below creates it for you with /skill-creator, Claude's built-in skill for making skills.
Once it's saved, Claude writes one detailed image prompt per slide.
Every prompt has the size, the exact text in quotes, hex colours and a list of things to avoid.
I ask for Slide 1 first, then the rest once the look is right.

Carousel image prompt skill

/skill-creator
Create a skill that turns my carousel copy into ChatGPT image
prompts, one slide at a time.
1. Every prompt starts with the canvas: PORTRAIT, 1080 x 1440,
   3:4. Never square, never 9:16.
2. From Slide 2 on, open with: This is Slide [N] of the same
   carousel. Match Slide 1's visual system.
3. Sections in order: BACKGROUND, HEADLINE, MAIN VISUAL,
   CALLOUTS, BOTTOM LEFT, AVOID.
4. Every visible word goes in quotes. Every colour gets a hex
   code. My palette: [your colours].
5. Main visual = real, nameable things. For app screens, use the
   real UI and labels. Never invent the output.
6. AVOID: at least 8 specific items, including 3D tilt,
   stretched screens and misspelled text.
7. End every prompt with: Ensure all visible text is correctly
   spelled. Generate and show me.
8. Write Slide 1 first and wait for my feedback before the rest.
9. Give me each prompt in a code block, ready to paste.

3

Step

Generate Slide 1 and lock the look

ChatGPT

I paste Slide 1's prompt into a new ChatGPT chat.
I get the headline, colours, fonts and layout right before moving on.

Cover slide, product on a laptop

This is Slide 1 (cover) of an Instagram carousel about [tool].
Create a PORTRAIT image. 1080 pixels wide by 1440 pixels tall.
3:4 vertical. Not square, not landscape, not 9:16.
SCENE: A dark, cinematic home office at night. A laptop on a dark
wooden desk. Soft city lights through the window behind. The only
strong light is the laptop screen glow.
LAPTOP SCREEN (the hero, sharp and readable): the real [tool] app
UI showing [what it's doing]. The [tool] logo clearly visible in
the app header. Nothing floats outside the screen.
TEXT (top third, on the dark background):
Line 1, bold white sans-serif: "[line 1]"
Line 2, larger, in [accent colour]: "[line 2]"
BOTTOM LEFT: "@[your_handle]"
AVOID: floating UI cards, 3D-tilted UI, other logos, misspelled
text, any orientation other than portrait 1080 x 1440.
Ensure all visible text is correctly spelled.
Final size: 1080 x 1440 px, 3:4 portrait.

4

Step

Generate the rest in the same chat

ChatGPT

Every slide stays in the same chat so ChatGPT keeps the style.
I generate one slide at a time and check the text before continuing.
If one detail is wrong, I fix only that instead of regenerating.

Result slide with callouts

This is Slide [number] of the same carousel. Match Slide 1's
visual system.
Create a PORTRAIT image. 1080 pixels wide by 1440 pixels tall.
BACKGROUND: flat [colour]. No texture, no gradient.
HEADLINE (top left): "[headline]"
MAIN VISUAL: a flat [tool] app window, perfectly flat, no 3D
tilt, rendered complete with no loading states. It shows:
[describe the real output].
CALLOUTS: [number] numbered cards around the window's edges, each
connected to one element by a thin 1px leader line ending in a
small dot. Leader lines never cross. Each card: small numbered
circle, bold 3-5 word title, one line of grey body text.
Callout 1 → [element]: "[title]" / "[body]"
Callout 2 → [element]: "[title]" / "[body]"
AVOID: loud ALL-CAPS labels, heavy boxes, crossing lines,
misspelled text.
Final size: 1080 x 1440 px, 3:4 portrait.

Next slide, same style

This is Slide [number] of the same carousel. Match Slide [X]'s
layout EXACTLY: same background, fonts, margins and spacing.
Only change: [what's different].
1080 x 1440 px, 3:4 portrait. Ensure all text is spelled correctly.

Fix one thing without regenerating

Keep everything exactly the same. Only change [the one thing].
Same size, same position. 1080 x 1440 portrait.

5

Step

Animate the cover (optional)

Codex

For some carousels, I ask Codex to turn the cover into a short 3:4 animation, using HyperFrames (see Reels, Make the b-roll).
Same 1080 x 1440 size as the other slides, then it holds on the final frame.

↑ Back to contents

3. LinkedIn and X posts

Claude

ChatGPT

Also

Codex

Written in my voice, with an image when the post needs one.

1

Step

Brainstorm, then draft

Claude

X

I use the same brainstorm prompt as reels, then pick one angle.
I have a separate writing skill for each: LinkedIn for finance leaders and builders, X shorter and punchier. Create yours with the "Create your own skill" prompt at the end of this guide.
No skill yet? Use the prompt below.

LinkedIn post

Write a LinkedIn post on [the angle I picked].
Audience: [who reads it].
1. The first line is the insight, not a setup.
2. Short paragraphs, one idea each. No hashtags in the body.
3. Use this real example or number: [paste it].
4. End with a question people can actually answer.
5. No em dashes. Never use: leverage, unlock, seamless,
   game-changer, delve, elevate.
Give me 2 versions: a full post and a 2-line take.

2

Step

Add one image

Claude

ChatGPT

Claude suggests quotes from the post and I pick one.
It writes the ChatGPT prompt, and ChatGPT makes the image.

LinkedIn post image

Give me a ChatGPT prompt for one image to go with this post.
1. LinkedIn carousel style, but a single image. 4:5 portrait.
2. Suggest 3 short quotes from the post first. Wait for me to
   pick one.
3. Put the quote in quotes, with the exact font size and hex
   colours.
4. Clean, minimal and readable on a phone.
5. End with: Ensure all visible text is correctly spelled.

3

Step

Track what's working

Cowork

Cowork reads my LinkedIn profile and recent posts through Claude in Chrome (Claude's browser extension, while I'm logged in) and builds a live dashboard.
I set up a scheduled task in Cowork that refreshes the numbers every week and suggests my next topics.

4

Step

Repurpose Threads posts for X

Codex

Threads

X

I've saved this as a skill in Codex. It opens my Threads in Chrome, reviews my recent posts and picks the strongest ones.
It rewrites them for X and gives me a ranked draft list.
Nothing is published unless I say so.

↑ Back to contents

4. Client work

Cowork

My client business runs on Cowork, which works directly with the files on my computer.

1

Step

Track leads and clients

Cowork

One Excel client list for clients, leads and brand collabs, with a live dashboard on top.
I just tell Cowork "add him as an active client" or ask "who are my open clients?"

Cowork, client tracker

Build a client tracker from [folder or sheet name].
1. One row per lead or client: name, company, source, stage,
   value, last contact, next step, due date.
2. Stages: [New lead / Call booked / Proposal sent / Won / Lost].
3. Dashboard on top: pipeline value by stage, count per stage,
   and every follow-up that's overdue highlighted.
4. Whenever I add new notes to the folder, update the tracker
   and tell me what changed.

2

Step

Turn notes into proposals

Cowork

I point it at my notes and a past proposal to match.
It drafts the proposal and a cover email, and flags anything missing.

Cowork, proposal from notes

Draft a proposal for [client name] using my notes in [folder].
1. Sections: the problem, what we'll do, deliverables, timeline,
   price, next steps.
2. Price: [your pricing or range]. Timeline: [weeks].
3. Match the format and tone of [past proposal file].
4. Also draft a short cover email to send with it.
5. Flag anything missing from my notes. Don't fill gaps with
   guesses.

3

Step

Turn meeting notes into actions

Cowork

Decisions, action items and open questions from every call.
It drafts the follow-up email and updates the tracker.

Cowork, meeting notes to actions

Here are my notes from the call with [client]: [paste notes or
attach transcript].
1. Decisions made.
2. Action items with owner and due date.
3. Open questions still unanswered.
4. Draft a follow-up email to [client] with the action items.
5. Update [client]'s row in my client tracker.

4

Step

Schedule weekly client updates

Cowork

Every Friday it drafts an update for each active client.
Drafts only. I review them before sending.

Cowork, weekly client update

Every [Friday at 5pm], for each active client in my tracker:
1. Draft a short update: what was done this week, what's next,
   anything blocked.
2. Pull details from [client folders / tracker notes] only.
3. Save each one as a draft in [Gmail / Outlook]. Don't send.
4. Give me one summary list of all drafts created.

5

Step

Email, files and browser tasks

Claude

Instagram

I connect Gmail, Microsoft 365 and Canva in Claude's connector settings so it can work with my source material.
For browser tasks, I install Claude in Chrome, Claude's browser extension, and use it for things like auditing an Instagram or LinkedIn profile.
I review anything that will be sent or posted before it goes out.

↑ Back to contents

5. Dashboards from data

Cowork

Copilot

Cowork for files on my computer. Copilot for files on SharePoint.

1

Step

Turn a folder of Excel files into a dashboard

Cowork

I point Cowork at the folder and ask it to combine the files.
It builds KPI cards, charts and filters from the source data.
I ask for a Refresh button to re-read the files and a check that totals match.
My own P&L and weekly content tracker run this way.

Cowork, folder to dashboard

Build a dashboard from the Excel files in [folder name].
1. Read every file and combine the data. Real numbers only, no
   placeholders.
2. Header bar with the title and entry count, plus a Sync button
   that re-reads the folder.
3. Filter row: [period toggle], [category dropdown], [month dropdown].
4. Six KPI cards with thin coloured left borders: [total income],
   [total expenses], [net], [margin %], [average monthly cost],
   [best month].
5. Charts: [income vs expenses by month], [split by category],
   [trend line].
6. Before finishing, check every total matches the source files
   and list any mismatch.

2

Step

Build an agent from a workbook

Copilot

I ask Copilot questions about my Excel data and the notes behind the numbers.
I build workbook agents that answer questions and produce dashboards checked against the source file.

Build an agent from a workbook

Create an agent that acts as a senior [role, e.g. FP&A analyst]
for my [file name] on SharePoint, with these capabilities:
1. Build a real HTML dashboard whenever asked, not a description
   of one: a dark sidebar with page navigation, a header with the
   source file name, a persistent filter bar, KPI cards with
   sparklines, and charts for [what matters to you].
2. Validate that dashboard totals match the source file before
   finalising, and flag any mismatch.
3. Answer direct questions about the data, like what's driving a
   variance or what changed month over month.
4. Explain variances using the notes in the source data, or say
   clearly when no explanation exists. Never guess.
Set up a suggested prompt for each of these four capabilities.

↑ Back to contents

6. Building and automating

Claude Code

Codex

Pages for the site, and the repetitive work on my Mac.

1

Step

Build guide pages

Claude Code

I give it the page content and point it at an existing page on my site to copy the layout from.
It builds the page, including collapsible prompts and copy buttons.
I check the local mobile preview before publishing.

Claude Code, build a guide page

Build a new guide page at /[slug].
1. Find my most recently built page and use it as the exact
   template. Reuse its header, footer, components and styles.
   No new colours or fonts.
2. No hero section. Title, then straight into Step 1.
3. Numbered steps, then collapsible prompt blocks (collapsed by
   default, one-click copy), then a feature card grid, then my
   existing CTA block.
4. Content: [paste your steps and prompts].
5. Mobile first. Run it locally, check a 390px viewport, and show
   me a screenshot.

2

Step

Automate repetitive tasks on my Mac

Codex

I give Codex permission to control my computer, then use it for repetitive tasks across apps.
Anything I repeat gets saved as a skill in Codex, the same way as in Claude.

↑ Back to contents

7. Learning from videos

Gemini

Gemini handles watching and transcribing.

1

Step

Summarise YouTube videos

Gemini

YouTube

I paste the YouTube link into Gemini.
I ask for the main point, key ideas, tools and examples, with timestamps.
I use the timestamps to find the parts worth watching in full.

YouTube breakdown

Summarise this video: [YouTube link]
1. The main point in two sentences.
2. Key ideas as a numbered list, with timestamps.
3. Every tool, number and example mentioned, exactly as stated.
4. The 2 or 3 moments worth watching in full, with timestamps.
5. One line on who this video is actually useful for.

2

Step

Transcribe videos

Gemini

I upload the video and ask for a word-for-word transcript with timestamps.
I also ask for a clean version without filler words, split into short paragraphs.

Transcript

Transcribe this video word for word, with timestamps every
[30] seconds. Then give me a clean version: filler words removed,
split into short paragraphs, speaker names if there's more than one.

↑ Back to contents

8. My 9 to 5

Copilot

Copilot runs my day job inside Microsoft 365.

1

Step

Recap meetings and draft follow-ups

Copilot

Decisions, action items, owners and due dates from every meeting.
It drafts the follow-up email from those details.

Meeting recap

Recap the [meeting name] meeting from [date].
1. Decisions made.
2. Action items with owners and due dates.
3. Open questions nobody answered.
4. Draft a follow-up email to attendees with the action items.

2

Step

Turn documents into decks

Copilot

I turn work documents into PowerPoint decks.
I restyle existing decks without changing their content.

Deck from a document

Create a [number]-slide PowerPoint from [file name].
1. Slide 1: title and one-line summary.
2. One idea per slide, max 3 bullets each.
3. Use a chart wherever there are numbers.
4. Final slide: next steps with owners.

Make any deck look designed

Apply this design method to the presentation I'm working on.
Keep my content, structure and slide count exactly as they are.
Only restyle the visuals.
Design method: [name, e.g. editorial / minimal / bold]
Visual direction: [e.g. warm cream background, one rust-red
accent, bold uppercase headlines, italic serif body text]
Palette: background [hex], accent [hex], deep accent [hex].
Typography: [headline font] for titles, [body font] for body text.
Density: [low / medium / high].
Use the design grammar (spacing, contrast, type hierarchy, colour
logic) consistently on every slide. Make it polished and
presentation-ready.

3

Step

Test new Copilot features

Copilot

I try each feature with my own documents, workbooks or meetings before posting about it.
I check the output against the source and note where it gets things wrong.

↑ Back to contents

Doing the same task every week? Turn it into a skill. Type /skill-creator in Claude, or just ask Claude to create a skill for it. Once it's saved, Claude follows your rules every time without you re-explaining them.

Create your own skill

/skill-creator
Create a skill for [the task you repeat, e.g. writing my weekly
client update].
1. Here is how I do it today: [paste your steps or an example].
2. Rules it must always follow: [tone, format, length, words to
   avoid].
3. Here are 2 examples of good output: [paste them].
4. It should trigger when I say things like: [e.g. "weekly update",
   "client update for X"].
Ask me anything you need before you build it.

↑ Back to contents