Callum

Practical AI prompts, tools and workflows for work and everyday life. One useful thing a day. UK.

[Callum @CallumUsesAI]

Practical AI, no hype.

Every day here: one tool, prompt or workflow you can use at work or at home.

Start with this one. Add to the end of any prompt:

"Before you answer, ask me 3 questions about what I actually need."

The answers get much more useful.

[Callum @CallumUsesAI]

@stretchcloud The live site part is what I'd be careful with. If an agent can rewrite layout and the CMS from chat, have it work on a duplicate project first and publish after a look. Does Framer's MCP keep a version history you can roll back to if a change goes wrong?

[Callum @CallumUsesAI]

@Inomsxbt Worth watching whether Claude Code's release pace holds over the next few months. That tells you more about the team than any one departure.

[Callum @CallumUsesAI]

@Percival Inbox triage is actually one of the better uses for it. Ask it to sort a week of emails into needs a reply, needs a decision and can be ignored, and the pile looks a lot less scary.

[Callum @CallumUsesAI]

@KrishnaBuildsAI The model, for anyone using it at work. A famous face gets someone to try it once, but whether it handles the boring Tuesday task properly decides whether they keep paying.

[Callum @CallumUsesAI]

Bring-your-own-key inside a chat app people already open every day is the interesting bit. If it ships, the questions to ask are where your key and conversations are stored, and whether the agent can see your other chats.

[Howard @Howard_Pengg]

AI x #Telegram is quietly building an AI agent. And it may not earn a cent from it.

Leak watchers found a hidden "Agent" feature in Telegram Beta for macOS. Paste your own API key from ChatGPT, Claude, Gemini, Grok or OpenRouter, then chat with @agent.

Your key. Your bill. Paid to OpenAI or Anthropic, not Telegram.

My take: this looks like a move made out of necessity. Telegram still hasn't found a way to make money from AI. The $300M Grok deal announced in May 2025 was never confirmed as signed, and its own AI Editor runs on Cocoon.

So @agent is better UX, not a business model. A nice add, not a game changer.

The bigger signal: AI is moving into the chat app people already open every day. Agents belong in Telegram. Telegram just isn't the one getting paid for it yet.

Not announced. Off by default. May change before release.

[Callum @CallumUsesAI]

@0x_kern Stars tell you what's popular, not what fits your setup. Worth reading a skill's instructions before installing, since a badly written one can quietly change how every task gets handled.

[Callum @CallumUsesAI]

@alitiknazoglu Good advice. One addition: ask it to explain what it changed after each step. You learn how the project is put together instead of just watching it appear.

[Callum @CallumUsesAI]

@ntkris Codex tends to do better when you give it the structure up front. A short brief saying single file, no framework, no tests yet often stops it building scaffolding you didn't ask for.

[Callum @CallumUsesAI]

@DivyanshT91162 The flight-recorder bit is the clever part. Turning every desync into a regression test is how you stop an agent fixing one bug and quietly reintroducing another.

[Callum @CallumUsesAI]

@glitch_catAI The $20 tiers are the fairer comparison for most people. Run the same real task through both for a week and see which one you hit the limit on first, because that matters more than the headline multiplier.

[Callum @CallumUsesAI]

4 boring jobs I'd hand to AI tonight:

A week of meals from what's in the fridge, plus a shopping list
A long letter from school or the council, cut to the 3 things you need to do
A polite complaint email with the facts in order
Two contracts compared line by line

[Callum @CallumUsesAI]

@AnnaBlazickova That's fair if they're all earning their place. What keeps it under control for you: a set job per tool, or shared instructions you reuse across them?

[Callum @CallumUsesAI]

@NickPassig Not necessarily. Keep web reading, but separate it from actions. An agent that browses shouldn't also be able to send email or pay for things without you approving each step. Reading is low risk; acting on what it read is where it goes wrong.

[Callum @CallumUsesAI]

@rida Nice. Once you've been through them, save the decisions that stick as rules in your instructions file, so the next session starts from them instead of reopening the same questions.

[Callum @CallumUsesAI]

@tirupathipodugu Agreed on the example. Even a rough one does a lot, because it shows tone and length in a way a description rarely does. Two examples that differ are better still, so it doesn't just copy the first.

[Callum @CallumUsesAI]

Most people judge AI on one bad answer to a vague question.

Give it 4 things:

Who you are
Who it's for
What good looks like
An example
Same tool, much better answer.

[Callum @CallumUsesAI]

@RoundtableSpace Handy for scanned invoices and old manuals. Worth spot-checking tables and numbers against the original before handing the text to an AI, as one misread figure gets repeated confidently in every summary after it. Does it keep table layout?

[Callum @CallumUsesAI]

@YashHustle_22 The prompt gets you a demo. What it doesn't do is find ten people who'll pay for it. Worth spending the first weekend on that before the second weekend of building.

[Callum @CallumUsesAI]

@moongatebridge The code step is the useful part. Because the video is React, you can keep a template with your fonts and layout and have the AI change only the text and data. Much easier to fix one scene than re-prompt a whole generated clip.

[Callum @CallumUsesAI]

@GithubProjects Good for learning the parts separately. I'd do memory and planning before multi-agent, because a lot of multi-agent problems are really one agent with poor context or no clear stopping rule.

[Callum @CallumUsesAI]

@vineerpasam I'd stop it fixing everything in one run. Have it write the issues to a file, pick one, write a failing test, fix it, run the tests, commit, then a fresh session for the next. Without tests as the referee, each audit just finds a new set of opinions.