I Don’t Pick Productivity Apps for Features Anymore

When Brooks looks at a new tool, he asks two questions. Can I get value from this quickly? Will it grow with me?

That’s his filter for shiny object syndrome. Two questions. I run the same test, and I add one more: does this thing actually talk to my AI?

Once you treat AI as a layer between you and your software, the feature list stops being the point. A complex task manager that traps your data is worse than a simpler one your agents can write to. I pick a task manager because it integrates with my AI agent platform, not because it has the longest checklist of bells and whistles.

Todoist is the clean example. It’s simpler than a lot of the “serious” task managers. Because it has a real API, an AI agent can ingest a meeting transcript, create the tasks, and put due dates on them. The manual admin work disappears.

That’s the Integration-First Tool Selection Framework. When you evaluate new productivity software, prioritize tools with robust APIs that plug into your AI backbone (Lindy, Zapier, whatever you already run) over tools that are merely feature-rich.

Audit your current stack for API friendliness. Open each app and ask: can an agent read from this and write to it? If the answer is no, that app cannot keep up with how you should be working.

Can you get value from it this week?

A powerful tool is useless if the learning curve is a cliff.

Brooks’s Easy On-Ramp Rule is simple. Use a simple tool for simple needs. If you already know you’ll need more power later, choose something with a low barrier that scales. Don’t start on the complex tool that demands mastery on day one.

There’s a second filter that sits right next to that rule. The 30-Minute Stickiness Test. The best predictor of whether you’ll stick with a new productivity app is your emotional reaction in the first 30 minutes. If the UI feels clunky or the workflow feels forced during the trial, you will abandon it. Don’t force yourself to get used to a tool that feels wrong immediately. Find one that feels intuitive from the start.

Does this sound familiar? You installed the popular app, hated the first session, and told yourself you’d grow into it. You didn’t. Stop forcing a tool you dislike.

Pick one heavyweight tool you’re struggling with. Replace it with a simpler alternative for 30 days and see if your week actually gets easier.

One backbone, not a pile of apps

Do you have three apps doing roughly the same job?

I don’t spread similar work across a pile of specialists. I consolidate around one backbone tool that does one thing exceptionally well and integrates with everything else. That cuts cognitive load and context switching. Name your backbone app. If you can’t, you don’t have one. You have a pile.

Same test for notes. A closed cloud platform owns your data. A local, file-based system like Obsidian lets you run local LLMs pre-trained on your own notes. You get the intelligence of a cloud AI with the privacy and control of local storage, plus the flexibility to hook it to any AI tool via MCP (Model Context Protocol) services.

Evaluate your note-taking app for two things: data ownership and API accessibility. If you can’t export the files and you can’t connect an agent, you’re locked in.

Agents are process design

Building AI agents is not a technical skill reserved for programmers. It’s a design skill. The coding is click-and-point or prompt-based now. The hard part is process mapping: if this happens, then do that. If you can map your manual workflow, you can build the agent to automate it.

Map one repetitive workflow on paper this week. That’s the whole homework.

Email is the obvious place to start. Don’t process messages one by one. Use an AI agent to scan your inbox on a schedule (hourly works). Give it your pre-defined guidelines, business context, and goals. It drafts responses. You review, approve, or tweak. Inbox becomes a quick approval task. You process it 10x faster, for a low monthly cost, and you don’t need to write code.

Use this as the template. Scan unread mail every hour. Draft replies from my guidelines and goals. Never send without my approval. Flag anything that needs a human decision. Then you sit down, approve the stack, and close it.

Right model, fewer clicks, talk instead of type

Don’t run every task through one general-purpose model. Curate a stack. Use a general model like ChatGPT as the daily driver, a technical model like Claude for coding and complex logic, and a creative model like Gemini for copywriting and image generation. Each model has specific strengths. Using the right one for the job gets you better results than forcing one tool to do everything.

Which one is your daily driver? Which ones are specialists? Write that down.

Two more setups that remove friction once the stack is in place.

A programmable button deck like a Stream Deck. Press one button and it opens a specific app, resizes windows to your preferred layout, and launches the associated tools (a terminal, an AI chat) at the same time. Starting a task goes from a pile of clicks to one physical action. Try a one-button workflow for your most common task.

And semantic dictation. Tools like Whisper Flow interpret intent instead of transcribing every “um.” Standard dictation captures corrections as extra words. Semantic dictation understands context. You say “three cheesecakes,” then correct it to “two,” and it writes two. Say “my home address” and it fills the stored address. Muscle-memory text expansion gets replaced by the tool understanding what you meant. Test one for a week.

Do this today

Don’t rebuild the whole stack tonight. Pick the one heavyweight app you already resent. Check whether it has an API your AI backbone can use. If it doesn’t, or if it failed the 30-minute test months ago and you’re still fighting it, replace it with a simpler, API-friendly alternative for the next 30 days.

One tool. Thirty days. Run that experiment.