Try itNov 13
The 4 steps:
1. Go to Meetings in Claude (claude.ai/meetings).
2. Connect your Google Calendar. An owner has to turn this on for your organization first.
3. Select a meeting on your calendar.
4. Click "Add Claude". Claude joins the call as a visible participant.
The links:
Claude Meetings: https://claude.ai/meetings
Anthropic's help article: https://support.claude.com/en/articles/18325154-use-claude-meetings
Launch film: https://www.youtube.com/watch?v=YUYeHMsqbmY
For your business: Worth trying if your team already meets in Google Meet on a Claude Team or Enterprise plan, because the notes, the transcript and the action items get written while you talk. It's in beta and works in Google Meet only for now. It transcribes the call, so tell everyone on it before you add Claude.
Try itNov 13
The nine steps:
1. Make a free Composio account at composio.dev. You don't need a credit card.
2. In Claude, open Settings, then Connectors.
3. Click Add custom connector.
4. Name it Composio and paste https://connect.composio.dev/mcp
5. Click Add, then Connect, and sign in to Composio.
6. Check that Composio shows in your connectors list with a CUSTOM badge.
7. Start a new chat and ask Claude to do something in one of your apps, like "summarize my last five Gmail emails."
8. The first time you use an app, Composio sends a sign-in link. Approve it once and it stays connected.
9. Turn off the single-app connectors Composio now covers, so they stop taking up room.
Using Claude Code instead? Paste this: Install the Composio CLI: curl -fsSL https://composio.dev/install | sh, then run composio login.
The links:
Composio: https://composio.dev
Setup docs: https://docs.composio.dev/docs/composio-connect
Pricing: https://composio.dev/pricing
For your business: Worth it if you've connected a lot of apps to Claude and your chats run out of room early. Claude Code already loads connector tools only when it needs them, so this matters most in the Claude app.
Try itNov 13
What's in the doc:
1. A niche sheet: 6 kinds of business and the step that slows each one down.
2. The first message, word for word.
3. Five call questions that find the money.
4. The pricing math, with a fill-in template.
5. The Claude prompt that builds the demo from the owner's own enquiries.
6. A 10 run scorecard, so no client sees a bad output.
7. Three objections and what to say back.
Every number in it is an example. Swap in the owner's real numbers.
For your business: Written for someone selling AI help to local businesses. If you own the business, the five call questions and the pricing math will still show you which step costs you the most.
Try itNov 11
The three steps:
1. Claim and verify. Go to Business Hub and claim your business. Once it's verified you can update your details and upload a product feed with prices and stock. Shopify stores can connect from Business Hub or the Shopify admin.
2. Open your dashboard. Business Hub shows how many times ChatGPT mentions you, your profile views, your website clicks, and the topics people ask about when your site comes up. The numbers only show once there are enough conversations.
3. Turn on your Business Agent. Set what it knows and how it answers, test it, then publish. Shoppers can open a chat from your ChatGPT profile and ask which product fits them.
The links:
Business Hub: https://chatgpt.com/business-hub
OpenAI's announcement: https://openai.com/index/introducing-businesses-chatgpt/
Learn more and claim: https://openai.com/business/businesses-in-chatgpt/
Who can use Business Agents: https://openai.com/policies/business-agents-policy/
For your business: Worth claiming now if your customers already ask ChatGPT where to buy what you sell. You give ChatGPT accurate details and you can see whether it mentions you. Business Agents are in beta for US ecommerce and online service businesses only, and OpenAI says ChatGPT's own answers stay independent of them.
Try itNov 10
This editor is fake. Claude Fable 5.2 wrote it as one web page in HTML and JavaScript, and nothing in it edits video. Ten labelled AI cursors cut a reel in one 29.4 second take, locked to the music track.
How it was built:
1. Claude Opus 5 built a first version that morning, with the editor layout and random motion. I switched to Fable 5.2 to rebuild the motion, and restructured my brief (Prompt 1) with Anthropic's guide to prompting Fable 5.2: an ownership sentence, a finish-the-whole-task line, named approvals and plain reports.
2. Fable 5.2 ran it in Claude Code with UltraCode on (/effort ultracode), It ran one workflow of six Fable 5.2 agents to pull the look, test effects and pitch the choreography, then built it.
3. It measured the track into 139 hits at 97.95 BPM and drew every frame from the audio time. The drop lands at 19.92 s, and every word in the final cut lands on a hit.
4. It tested itself in headless Chrome: two full takes of 1,859 and 1,860 frames, 16.67 ms average frame time, zero frames over 20 ms and zero console errors.
5. I filmed my monitor with my phone in one 31.5 second take.
Prompt 1, my brief, shortened (fill in the brackets):
"You are the motion designer and the engineer on this project. This codebase is yours: what it's for and how it's built are your call.
What we're making: a fake video editor in HTML and JavaScript that fills my screen ([width] x [height] px). I'll film the screen with my phone and post it as a reel. Ten AI agents with labelled cursors cut a reel together in one [length] second take, locked to [song file]. Every change on screen happens because a cursor clicks on a beat. Nothing in it edits real video.
The sound: measure the track first (BPM, every hit, the breakdown, the drop) and save the hits to a cues file. The whole picture is a function of audio time. Add a t setting to the address (for example t=12.5) that renders that exact frame with the audio paused. Big moves land within one frame of a measured hit. Hold back in the breakdown, hit hard on the drop and end on one clean hero frame.
Hard limits: it opens by double-click in Chrome with no server and no build step. Make every graphic in code, with no AI images, stock, real people, real logos or invented stats. Plain English on screen. At the drop and the last frame, everything that matters sits in the centre 9:16 column so the video crops to vertical.
Needs no approval: reading and writing in this folder, and running headless Chrome, ffmpeg, python3 and node for screenshots and measurements, including retries.
Never on this job: posting, pushing, deploying, sending, spending money, installing packages or deleting files.
Finish the whole task. Don't stop while you still owe work this brief asks for. Work through errors, timeouts and flaky runs yourself. Stop only for a real blocker someone set on purpose, and say plainly what you found.
Done means the take plays start to finish with zero console errors, every scripted change lands on its hit, and the frames look designed at full size and in the 9:16 crop.
Report for someone who hasn't seen your tool calls: key points first, short sentences, what you built, what you checked and how, and what you didn't check."
Where Higgsfield fits:
Higgsfield made one test still for this build, and it isn't in the posted video. To fill the editor's monitor after the drop, I'd connect the Higgsfield MCP to Claude and have its Katana workflow make an 11 second clip from a reference video. Seedance 2.5 costs 7 credits a second at 720p and 16 at 1080p (Higgsfield's free price check, Nov 10), so 11 seconds at 720p comes to about 77 credits. That's an estimate.
Prompt 2, with the Higgsfield MCP connected:
"This one clip is the only exception to the brief's no-AI-images rule. Everything else stays drawn in code. Use Higgsfield's Katana workflow to make an 11 second black and white clip from this reference video: [link to a video you own or have the rights to use]. Follow its cuts and camera moves. The subject is an invented, faceless performer, never a real person. Use Seedance 2.5 at 720p, sized to fit my editor's program monitor ([aspect ratio]). Before any paid job, run the free price check, show me the credits and my balance, and wait for my OK. Never buy credits or change my plan. When it's done, save the clip in the editor's folder, play it in the program monitor from the drop to the last frame on the audio clock, and log the job id and the credits spent."
For your business: Use the same brief for your own product demo or launch video. Keep its parts: who owns the work, what done means, what needs no approval and what's off limits. The 77 credit Higgsfield figure is an estimate, so have Claude show you the price before any paid job runs.
Try itNov 9
Setup:
1. Paste this into Claude: https://github.com/nessalazne/linkedin-agent-skill Install this skill, then confirm /li-post works.
2. Fill in your voice file. Copy templates/voice.md to ~/.claude/linkedin/voice.md, or paste three of your own posts into Claude and say "write my voice.md from these". Every skill reads that file, and the drafts sound generic without it.
3. To post from Claude, put BLOTATO_API_KEY and BLOTATO_ACCOUNT_LINKEDIN in ~/.claude/linkedin/.env. Then run python3 publish.py --check to confirm the account is LinkedIn.
The twelve from the reel:
/li-post writes a post from 21 hook formulas.
/li-comment comments on other people's posts.
/li-reply handles the thread under your own post.
/li-profile scores your profile out of 100.
/li-plan plans the week and picks 10 people to engage with.
/li-human strips the AI tells and scores the draft.
/li-carousel builds document posts.
/li-repurpose turns one video or newsletter into a week of posts.
/li-dm writes the invite note and follow-ups.
/li-inbox sorts your inbox.
/li-audit ranks what you already published.
/li-publish posts or schedules through Blotato.
The repo has added three more since (/li-interview, /li-hooks, /li-engagers).
For your business: Good if LinkedIn brings you clients and you keep skipping it. Read every draft first: /li-publish shows the exact text and waits for you to type publish, and the other skills never touch your account.
