Les Thumbnails d'Allan

View details

I cancelled GrokBot. Then I went back.

Having the right capabilities and actually getting work off your plate is not the same thing

Test one real workflow and assess how much coordination the assistant carries without repeated prompting.

View details

A one-day trial. A lasting workflow upgrade.

Trying GrokBot helped me get more from the tools I already had.

Trying a new tool can reveal better ways to use the tools you already have without migrating.

View details

How to create a self-improvement loop for your AI

Don’t correct the same AI mistake every month

Turn reusable feedback into a corrections file the agent reads before the next batch.

View details

Get the keys to your app

Four things to clarify, plus a prompt to help you find them in your project.

Confirm business control of the source, accounts, connections and costs before completing an app handover.

View details

Computer use is the future (bring a credit card)

A short documentary from a night of Grok, Codex, and burning through the meter.

Computer-use workflows that let agents coordinate with each other can unlock daily leverage, but they burn allowance so fast that plan upgrades alone may not keep up.

View details

Give your app a CMS without starting from scratch

Collections, admin screens, and approvals for people who aren’t technical

When non-technical people need to update content, use a ready admin with collections and approvals instead of building a CMS from scratch or forcing them into developer workflows.

View details

Your app can scale. Does the model?

A platform can be right for validation and need another look when a larger customer arrives.

Evaluate platform costs against the growth roadmap before committing to an approach; include the cost and timeline of a later rebuild.

View details

Save time before building the feature, too

Let AI investigate and prepare the task, so you can focus on deciding what to do.

Let AI investigate customer requests and prepare reviewable tasks before implementation, reserving founder time for decisions.

View details

“Can we add a delete button?”

Removing something from the active list and erasing its history are different product decisions.

Before adding Delete, define what should disappear, what history must remain, and what retained records should show afterward.

View details

Let AI fight with your page builder

We brought an existing design into Kajabi without recreating every detail by hand.

When a page builder makes a known design tedious to reproduce, let an AI agent operate the existing interface after you approve the structure and visual target.

View details

Watch the first run

You don’t want to supervise your agents forever. You also need to know when you can trust them to do the job properly

Watch an agent's early autonomous runs closely enough to correct what done means, then rely on trustworthy evidence instead of permanent supervision.

View details

1 hour 19 minutes without Goal mode

The ceiling is higher than you think, and you don’t need special techniques to find it

Before assuming an AI agent has reached its limit, ask it to take the work one step further and identify only what truly requires you.

View details

Your agent needs a sandbox

Testing gets complicated when your agent works across business apps.

Test one agent workflow with realistic fake records, redirected emails and resettable actions before allowing it to touch live business systems.

View details

What could go wrong?

A question that belongs in the plan, not after implementation.

Ask an agent to review unwanted side effects before implementation, then make the remaining risks conscious decisions appropriate to the context.

View details

Less maintenance. Still not set-and-forget.

The platform maintains the software. You still have to respond when a connected service changes.

No-code platforms reduce technical maintenance, but every business-critical automation still needs a named internal owner who reads service notices and can update and test the workflow.

View details

Migrating from Microsoft to Google got easier

Google Workspace can now carry over more of the files, permissions and conversations that admins previously had to rebuild.

Built-in migration tools reduce the administrative effort and uncertainty of moving files, permissions, and conversation history from Microsoft to Google Workspace.

View details

Make your AI images editable

Change one detail without regenerating the full image.

Use Magic Layers to edit individual objects in an AI-generated image without regenerating the entire composition.

View details

Avoid the insider job

The dangerous gap is between what an agent can do and what you meant.

Narrow an agent's assignment before granting it access, and require approval for consequential actions.

View details

AI automation no longer waits for the clock

Event triggers are replacing the hourly check.

Use schedules for time-based work, deterministic automation when an event always means the same action, and event-triggered AI when the response requires context and judgment.

View details

Want to replace your tech team with AI?

Meta tried. And failed.

AI-generated implementation only becomes productivity when technical judgment turns it into a safe, useful and maintainable product.

View details

Is it worth owning the whole stack?

The best solution may be the one you don’t build from scratch.

Before building a custom app, decide whether its specific value justifies owning the infrastructure, updates, security patches, and inevitable surprises underneath it.

View details

AI is learning to get to the point

Finally discovering that the status meeting could have been an email.

The best AI output reduces reading and fatigue without hiding verification, risks or decisions.

View details

Use the technology you sell

Analytics can show you what happened, not how it felt.

Use your product to feel the friction, then confirm the problem with customer evidence before building.

View details

Claude’s watermark “mystery” is solved

Anthropic named the technology, and documented why it still isn’t proof.

A watermark can suggest Claude was involved, but short samples, factual text, light edits and full rewrites limit what detection proves.

View details

Save your humans’ time in the loop

Human oversight should consume judgment, not time spent discovering what needs judgment.

Manual AI handoffs should reach the right experienced person with a clear decision and enough context to act without investigating.

View details

Make lets AI build without hiding the workflow

Maia may be the ideal compromise between ease and control.

AI-assisted building is most useful when the result remains visible, inspectable, and editable.

View details

Ask PAM if AI needs you

Not every AI output deserves your attention.

Route AI outputs through Pass, Automate, or Manual so human judgment is reserved for ambiguous, consequential, or hard-to-reverse decisions.

View details

Company data > Customer data

Google bid millions for the emails and Teams messages that could teach AI how work really gets done.

Internal communications may be valuable AI training data because they reveal the unofficial human patterns behind how work gets done.

View details

Fine. I'll update Zoom.

Software updates are annoying, until you learn what they fixed.

Install security updates promptly, especially on tools used for client calls and everyday business operations.

View details

Now Anthropic wants to watermark AI writing…

We keep trying to prove AI was involved. It is still measuring the wrong thing.

Do not outsource judgment to AI—or to tools that claim to prove AI involvement; review the quality and consequences of the work itself.

View details

Your AI automation needs a fallback

The happy path is never the whole workflow.

Design the fallback, independent confirmation, and human takeover point before an important automation becomes operational.

View details

Opera might actually be better

Chrome didn’t win me back. It just made returning easy.

The most feature-rich tool does not always win; familiarity and integration can make a default easier to return to.

View details

Give the new idea its own path

When one discussion produces two pieces of work, a fork gives both the context without mixing them in the same thread.

Fork a conversation when shared context needs to produce a separate persistent outcome.

View details

Not every question belongs in your AI conversation

Side chats give you clarification without cluttering the work that follows.

Use the main conversation for questions that change the work; use a side chat for disposable clarification.

View details

Chrome has all this?!

Sometimes the tool isn’t outdated. Your understanding of it is.

Before replacing familiar software, check whether your understanding of it is outdated.

View details

Goal mode should not come first

OpenAI is prioritizing autonomy before alignment.

Use Goal mode only after the plan and verifiable finish line are settled.

View details

Any employee can teach AI a workflow

Instead of writing the instructions, they can show the agent how the work gets done.

Employees can teach AI repeatable workflows by demonstrating them, but recorded skills should be reviewed for security, bad habits, and accuracy before reuse.

View details

Your legacy software might finally have an exit

Computer use gives AI agents access to old systems that have no API, connector or useful export.

Computer use can make one-time migrations from closed legacy systems economically viable by letting agents work through the existing interface under human review.

View details

The first run costs time. The second one gives it back.

The frustrating setup becomes the reusable operational knowledge

The first AI workflow run may cost more time than it saves, but preserving its instructions and corrections lets later runs compound the return.

View details

Gamma beat my AI slide system. I’m keeping both.

Buy the result you need today while building the option you may want tomorrow.

Use the specialist tool that wins today while feeding its approved output back into the local system you want to improve for tomorrow.

View details

Use the most powerful model you can afford

The cheaper model costs more once you factor in your time.

Use the most capable model you can afford when the task is building or debugging; token savings do not justify losing your time.

View details

Don’t let AI turn your job into endless review

The more execution you delegate, the more your job becomes deciding what moves forward.

Define the return checklist before delegating so AI output comes back ready for a focused decision instead of a full reconstruction.

View details

Turn your subject into slides that feel like you

First get the story right. Then let it fill the template.

Approve the copy before AI fills a reusable slide template, then save the feedback so the next deck starts ahead.

View details

Build the template before the deck

Your old slides are more useful than your next prompt.

The best way to make AI presentations feel like yours is to turn your existing visual choices into an approved reference template.

View details

It looks done sooner

AI can make a feature feel finished long before it is ready to rely on.

A convincing interface can look finished long before it has earned the trust required to ship.

View details

What’s complex now?

AI didn’t just speed up software work. It scrambled our sense of what should be difficult.

AI makes the first build faster, but it does not make difficulty predictable.

View details

Start with the task you hate

Your first AI agent is not meant to maximize revenue, but to make you want to build the next one

Start your first AI agent with the task you most want to stop doing; it gives you a low-risk way to learn delegation before chasing revenue.

View details

Your subscriptions bill might be hiding duplicates

A quick way to spot overlapping tools before your next renewal.

Grouping subscriptions by the main job they perform reveals overlapping tools that may be worth downgrading or consolidating.

View details

Describe work, not software

Before asking AI for a screen, button, or automation, explain what needs to happen.

Describe the work and desired outcome before prescribing the screen, button, or automation.

View details

Your future AI infrastructure might not be a server room

Open models could soon let businesses run most of their AI internally

A single high-spec desktop could soon handle most private business AI workflows without cloud infrastructure or usage-based pricing.

View details

A feature nobody notices is not fully shipped

Products keep evolving. What users believe they can do often stays frozen in time.

A useful feature cannot improve the customer experience until the people it helps know it exists and understand what they can do with it.

View details

Small companies might discover AI's limits the expensive way

AI makes software cheaper to build. It does not make fragile systems cheaper to own.

Speed becomes a liability when AI-built systems outgrow the infrastructure and ownership behind them.

View details

Your community is not your company chat

The familiar choice may be working against what you are trying to build

For a cohort or client community, choose a chat tool whose history and pricing model let the value of member conversations compound.

View details

Fable won. Look again.

Before switching models, look beyond who supposedly won the comparison.

Model comparisons need context: tier, speed, cost, and resource use matter more than who produced the prettiest result.

View details

The browser is becoming a commodity

What to prioritize when every browser starts offering the same features.

Choose digital tools by where your work persists, not only by the features on screen.

View details

Your Demo data might be lying to you

Clean demo prove the idea can work. Does not make it complete.

A POC is not ready until messy inputs reveal both where it fails and how the system responds.

View details

A determined user always find a way

Don’t mistake client-side friction for security.

If users can view data, assume they can capture it; protect sensitive information by limiting access, not by disabling browser interactions.

View details

OpenAI just made switching from Claude an easy path

GPT 5.6 is a model launch. But the product decisions around it look like a deliberate effort to remove friction for people curious about ChatGPT agents.

When a competitor has already shaped user expectations, reducing switching friction can be more persuasive than trying to look completely different.

View details

Not every App needs an interface

Some workflows are designed to be operated by agents

Recurring workflows can become small agent-operated systems: more durable than prompts, lighter than traditional apps.

View details

Skills are better with tools

Instructions help the agent know what to do. Scripts help it do it the same way every time.

Skills become more useful when they package repeatable scripts and tools, not only instructions.

View details

There is nothing magical about Skills

They are to AI agents what recipes are to cooks.

Save the process that worked as a skill instead of re-explaining it next time.

View details

AI can help you think before it helps you write

Before asking for output, make it interview you.

AI becomes more useful when it interviews you before it writes for you.

View details

What Goal mode is for

A new agent mode for giving AI a finish line instead of another prompt.

Goal mode works best when the agent has an independently verifiable finish line.

View details

AI unlocks your true potential

Not because it lets you skip the work. Because it helps the work reach the level it was aiming for.

AI helps people who already have taste, intent, and persistence reach a level of quality that used to require a team or budget.

View details

The tiny reasons we stick with an App

Sometimes the winning feature is not on the spec sheet.

When apps reach feature parity, small UX choices and taste decide what people keep using.

View details

You can vibe-code Mac apps now

But Glaze's real trick is making these apps easier to launch, share, and install.

Glaze is interesting because it pairs AI app building with a runtime and store that make Mac apps easier to launch, share, and install.

View details

Stop copy-pasting into ChatGPT/Claude

Let AI work where your files already live

Connecting Drive to an agent closes the loop: AI can read the source, discuss the change, and write back instead of making you shuttle files.

View details

Skip Starter, get at least Standard for Google Workspace

10x the value for twice the price of Starter

Business Standard is where Google Workspace starts replacing separate subscriptions instead of only providing a business inbox.

View details

The first tool I add after going live

The best AI-first tools do not add complexity. They remove the work between question, insight, and action.

After launch, founders need tools that reduce the distance from product question to useful action.

View details

Never get attached to a specific technology

The tools that work today can still become tomorrow’s migration project.

Keep your systems stable, but stay ready to migrate when a critical tool changes or disappears.

View details

Which AI level does this job need?

The useful move is choosing the lowest level that removes the bottleneck.

Choose the AI level that fits the job instead of using chat or agents for everything.

View details

The AI ladder

Before chasing better models, figure out which level of AI your work actually needs.

Founders get better leverage by matching the AI level to the bottleneck instead of chasing the most advanced tool.

View details

Before building, define who

The role map comes first: who uses it, who maintains it, who approves it, and who owns it.

Start app projects by mapping who uses, maintains, approves, exports, fixes, and owns the workflow before designing screens.

View details

Learn as you delegate

Delegating saves time today. A few extra questions are what increase your leverage tomorrow.

Use every completed AI task to improve the next brief, not just accept the current output.

View details

AI can build the integration. Should you still use Zapier?

The real question is not whether DIY is possible. It is how much workflow you want to maintain afterward.

AI coding agents can build integrations, but they do not remove the setup, security, testing, edge-case, and maintenance ownership that comes afterward.

View details

What bad looks like is part of the brief

Showing the anti-pattern can be as useful as describing the target state.

Tell AI what to avoid so the useful format becomes a constraint, not a wish.

View details

Amazon has entered the personal agent room

Amazon Quick looked interesting for ten minutes. My existing context looked more valuable.

New AI tools need to justify the cost of leaving the context your current tools already have.

View details

How daily writing accidentally rebuilt my Instagram

Once AI joins your process, one project turns into the next, and the next…

AI turns one small habit into adjacent projects when it becomes part of the thinking process.

View details

Share the "Good to know" with your agent

When it feels barely relevant, it might actually be a key constraint

Small context details can change the architecture an AI agent designs.

View details

Make your AI setup portable

Computers became interchangeable once the work moved to the cloud. AI models are heading the same way.

Keep your AI setup portable so model changes feel like swapping machines, not starting over.

View details

ChatGPT has eyes on becoming a Super App

Still on a quest to become your de facto personal agent, one feature at a time

Small convenience features train users to make AI part of their daily default routine.

View details

Just because "AI can build it" does not mean you should

The real skill is knowing when to build, when to buy, and when the default tool is already enough

Build only when the workflow is specific, repeated, valuable, and still underserved by existing tools.

View details

Models can disappear. Your work should not.

The latest AI model drama is another reminder that resilience matters more than model hype.

Design AI workflows so a model outage or product decision does not stop the work.

View details

When to use Deep Research instead of Chat

Some questions need more than a reply

Use Deep Research when a question needs source-backed exploration, not just a quick chat answer.

View details

Can another agent pick up the work?

The real test of whether your AI workflow is resilient.

An AI workflow is resilient only when another agent can pick up the work from files, plans, and written context.

View details

How I burned through my AI allowance in 90 minutes

The agents weren't building. They were reacting.

Structure feedback into specs before dispatching agents, or AI work turns into reactive token churn.

View details

Your App probably doesn't need the App Store

Start web-first, then build native around the flows your users prove are worth it.

Start web-first until real usage shows which mobile flows deserve native investment.

View details

Apple arrives late to the AI game, but with a real advantage

Not the biggest model. The deepest access to the user.

Apple's AI advantage may come from owning the personal surface, not from winning benchmark comparisons.

View details

AI can know the answer and still miss the move

The agent may have the knowledge, the feature, and the capability. Expertise is still knowing when to ask for it.

AI can hold the right capability and still miss the move; expertise is knowing what to ask for and when to use it.

View details

Subagents are not just fast mode

The real gain is not parallel work. It is better delegation.

Subagents improve quality when broad work is split into meaningful specialties, not just when work runs in parallel.

View details

The answer is still to try it

Benedict Evans had a useful take on the AI question everyone keeps asking.

You do not get better at AI by watching from the outside. You get better by trying it on real problems.

View details

Not keeping up with AI? Be more curious

The people getting better are not watching every demo. They are testing AI against their own use cases.

Curiosity compounds when you test AI against your own projects.

View details

The AI agent you need might already exist in the tools you use

I used Amazon’s Rufus to turn a book search into a simple lead qualification workflow

The fastest AI workflow can be the agent already sitting inside the product.

View details

Your agent wanted to follow you…

It does now.

Agent work no longer has to stop when you leave your desk.

View details

Using AI every day but not getting better yourself?

You have to go beyond generation and delegation.

AI gets more valuable when you use it to learn, not only delegate.

View details

When You Need to Get Out of Chat

The moment ChatGPT should become Codex

Some tasks need an agent that can work with files and verify output.

View details

The real AI skill is resilience

Google’s latest AI editor update is a reminder that your workflow cannot depend on any single vendor.

Your workflow cannot depend too much on any single vendor.

View details

Reach > Ego

The iPhone moment at Google I/O says more about Gemini’s ambition than Pixel’s importance

Distribution often matters more than ecosystem pride.

View details

Vibe coding for e-commerce

Faster storefront builds, different tradeoffs

AI storefront building is faster, but the tradeoffs still matter.

View details

The most underrated app on your iPhone

Why Apple Reminders still beats most todo apps for iPhone users

The default app can win because it owns the ecosystem integration.

View details

Why Google might become the biggest AI provider

Winning the AI battle might have less to do with model quality than people think

Winning AI may depend less on model quality than reach.

View details

AI done right: the case of video editing

A masterclass from Riverside

AI works best when it helps exactly where a human gets stuck.

View details

When a Book becomes an App

How a project stayed with me and led somewhere new

A written concept can become an interactive experience.

View details

I hired a personal painter

Turns out she lives in my sidebar

Good visual generation depends on brief, context, and iteration.

View details

How to keep up with AI advancements?

Spoiler: You don't.

You do not keep up by watching everything. You keep up by tinkering.

View details

I ignored my own advice about AI workflows

A small limit hit turned into a surprisingly useful reminder

Do not let one generous tool make your workflow fragile.

View details

Your agent wants to follow you

Digital Nomad is becoming a whole new, more powerful thing

The agent session becomes more useful when it follows you.

View details

Remember when Wordpress was the default?

How AI tools are solving pain points we accepted for years

AI tools can remove pain points we accepted as normal.

View details

Context is the brief

More real-world examples of why vague AI briefs still bite

Vague instructions are risky when agents act inside real projects.

View details

Full-stack freedom has a price

Shipping the first release is one thing. Living with the build is another.

Shipping is easier with agents, but ownership still has a cost.

View details

For once, small businesses have the advantage

Don't waste the window

Small businesses can move faster while the AI window is still open.

View details

Get comfortable switching models

The best AI workflows are built on adaptation, not allegiance.

The best AI workflow is using the strongest tool available, then switching when the lead changes.

View details

The Developer (as we know it) is dead. Long live the Tech PM!

Why AI may not kill software careers — just promote the best developers into a new role.

AI shifts more software value toward product judgment and orchestration.

View details

The Claude Army is mainly paying for ego, not output

Most AI work is repetitive, not profound. Use enough model, not maximum model.

Use the model that fits the job instead of paying for ego.

View details

ChatGPT Pulse: OpenAI’s first step toward personal agents

When AI stops waiting for prompts and starts inferring relevant recommendations from past conversations

Personal agents begin when AI stops waiting for explicit prompts.

View details

Context switching is the new multitasking

In an AI-assisted workflow, the real drag isn’t doing the work — it’s reloading your brain between tasks.

In AI workflows, the bottleneck is often reloading context between tasks.

View details

Better AI outputs start with better inputs: a real-world example

View details

Garbage in, Garbage out...

Clarity in → Quality out

The quality of AI output depends on the clarity of the input.

View details

Gemini for Businesses just got smarter

An even more seamless experience for Google Workspace users... almost

Workspace AI gets stronger when it can use the context already inside the tools.

View details

It's all about the memory...

Why your favorite AI might just be the one that knows you best.

AI agents need durable context, not just code execution.

View details

Claude Design looks great. The comparison gets awkward.

A look at what it produced, how it compares with Lovable, Gemini, and Codex, and why the bill matters more than the hype.

AI design tools should be compared by output, workflow, and cost.

View details

Use Claude Code and Codex for more than Code

They’re great for writing, strategy, and ops too!

Coding agents are useful beyond code when they can carry context into writing, strategy, and ops.

View details

The problem with vibe coding is not the coding

A new debate has started around vibe coding.

AI can make building easier, but expertise is still needed before people rely on it.

View details

What would you use a personal agent for?

In yesterday’s post I talked about the benefits of personal agents.

Start personal-agent delegation with one small, repetitive, annoying task that does not really require you.

View details

Why your next upgrade is not an App, it’s a Personal Agent

The next interface shift is from operating apps yourself to briefing the agent that orchestrates them.

View details

Your Product Owner is now your best Tech Lead

A Product Owner can become the technical lead when clear specs guide AI agents.

View details

How to (really) use Plan Mode

Spend some time before, or waste more time after, your pick.

Plan mode works best when the agent clarifies assumptions, drafts the implementation plan, and waits for approval before execution.

View details

"AI can make mistakes. Check important info"

Sometimes with itself...

AI can be useful and still wrong; check important outputs before trusting them.

View details

AI is an Engine. You’re still the driver.

AI can provide the horsepower, but the human still chooses the direction.

View details

On working with Jimmy, Cody, and Clo

Yes, I give my agents nicknames

Strategic patience can look foolish until the ecosystem advantage becomes visible.

View details

The “Best Model” is a distraction

Stop picking. Start combining.

Stop chasing the single best model; combine multiple systems for better angles and leverage.

View details

How to get 60% off your Claude Max subscription 👀

Use Claude for high-level thinking and Codex for execution capacity instead of paying for one expensive bottleneck.

View details

Share this with everyone you know

Gmail dot variants and + aliases let one inbox behave like many dedicated addresses.

View details

Treat AI like an Intern

Very knowledgeable, very capable, but an intern still.

AI gets more useful when you onboard it with context, examples, constraints, and feedback.

View details

After Vibe coding, meet Vibe design

Turns out “I’m not creative” was just… a tools problem.

AI design tools turn taste and direction into leverage when you can steer the workflow.

View details

Embrace AI and start building tools for yourself

Just start!

The fastest way to understand AI is to use it to build small tools around your own problems.

View details

After Your Custom GPT Knows You, Here's What's Next

From Context to Execution: Two Real Workflows

Custom GPTs become execution engines when you teach them your process, reference work, and exact output format.

View details

Why I Pay for Google Workspace, Even Though Gmail is Free

Or when the grown-up version of the tools you're already using makes sense.

Pay for Workspace when custom email, shared ownership, storage, privacy, or bundled tools start mattering more than a free inbox.

View details

I talk more than I type now (and you should too)

One tool, one keystroke, every app. Super Whisper turns messy speech into clean text.

Talking to AI removes friction when the assistant can turn messy speech into usable work.

View details

Here's your ChatGPT problem

The quality of your ChatGPT response depends on the instructions you give upfront.

Reusable upfront context makes ChatGPT outputs better without rebuilding the prompt every time.

View details

I switched from Dia to ChatGPT Atlas...

Trying to find the AI browser most suitable to me

For founders, context-aware browsing only becomes powerful when the conversation remains persistent and easy to revisit across devices.