I Used to Have 7 Half-Done Projects. AI Gave Me 87.

Before AI, I had 7 half-done projects sitting around. After AI, I have 87.

Read that again. Faster building did not make me more finished. It made the pile bigger.

Here’s the thing. AI lets you build automations and systems faster than you can comprehend them. You shovel a task at the model, it hands you a workflow, and you move on. You never sit with the process long enough to understand it.

That’s the knowledge debt trap. Systems pile up. Understanding does not. And if you don’t understand the underlying process, you can’t tell whether the automation even belongs there.

Does this sound familiar to you?

Do the task three times before you automate it

Don’t build a system, a workflow, or an AI automation until you’ve performed the task manually at least three times.

The first few passes reveal the real roadblocks. They show you missing features and actual requirements that imaginary systems miss. You find out the process might not even be needed. Or that it’s easier to do by hand.

This runs against the hustle of immediate AI adoption. A task repeats, and the instinct is to automate it tonight. Slow down. Identify one recurring task you’ve been tempted to automate and commit to doing it manually three more times before you build anything.

Make a crappy first draft and keep going

Anne Lamott wrote about the crappy first draft in Bird by Bird. You can always refine later. You can’t refine a blank page.

Same principle for AI workflows. Instead of over-optimizing prompts and systems before you start, generate a crappy first draft with the AI. Over-planning and tool-switching love a blank page. A rough draft gives you something to iterate on.

This pairs with doing it three times. First draft. Then another pass. Then another. When you start a new project, use AI to generate a rough, unpolished draft immediately. Don’t wait for the perfect prompt.

Ignore the “would you also like to” loop

Have you noticed what AI tools do the second you finish something? They ask, “Would you like to also do X?”

That’s the scroll trap. Modern AI tools mirror social media addiction. You complete the task, then you spend half an hour refining a result you didn’t need. Endless optimization loops. You sat down to finish one thing and stayed to polish a thing that was already done.

Brooks has a rule for this. Keep the desired result front and center. If the AI suggestion doesn’t directly serve the immediate goal, ignore it and move on.

Treat that as digital hygiene. Next time you use an AI tool, consciously resist the follow-up suggestions unless they directly serve your current task.

Stop pushing information nobody pulls

Automated information pushed to you is easy to ignore. Daily digests. Auto-emails. The fire hose.

Brooks argues that pulled information, the kind you actively seek out, is more likely to get consumed.

So audit your current automated digests. If you don’t read them, delete them. If you want to keep one, reframe it as a suggested experiment rather than a data dump. One experiment to try this week. That framing is what makes a push worth opening.

One tweak a week beats a pile of bookmarks

You know the other pile. Bookmarks from X. Videos from YouTube. Saves you never open again.

Turn consumption into action. Create a weekly digest of those bookmarks that synthesizes the content and suggests one specific experiment to run. The LLM does the heavy lifting of synthesis. You pick the action.

That’s the one tweak a week habit. You implement at least one new idea per week. Act on one idea. Leave the rest of the pile alone.

If a full bookmark digest feels like a project, start lower. Manually dump this week’s meeting transcripts into an LLM and ask, “What should I do next week?”

Make docs humans can scan

Markdown is great for AI input. It’s poor for human scanning.

HTML pages let you use rich visuals. Tables. Diagrams. Flowcharts. Documentation that’s easier to read and act on.

I built a Claude Skill that auto-selects the best visual type for meeting prep and system docs. Venn, flowchart, timeline. It turns static notes into interactive, scannable artifacts.

You can try a smaller version today. Take a recent meeting transcript, give it to an LLM, and prompt for visuals and tables instead of plain text. One HTML artifact. If you’re not ready for HTML, use the Marked app to view the markdown.

Review the week so you can understand what you built

The 87 half-done projects are what knowledge debt looks like in a list.

The solution I’ve been using is a Weekly Retrospective automation. It reviews the week’s work, surfaces improvements, and delivers that as a digestible HTML artifact. That’s how I close the gap between building and understanding.

You can start simpler. Set up a weekly review, manual or automated, and audit your half-done projects. Look at what you actually comprehended this week versus what you shoveled to an AI.

Then do this one thing. Pick one recurring task you’ve been tempted to automate. Do it manually three more times before you build a system for it.