The Tech LeverageBuild with AI · Audit the risk Get The Brief

Case 015 · Build Lab

Using AI All Day and Still Behind? I Audited a Week of My Own Chats

I audited a week of my own AI sessions and found five bad-boss habits. Each has an old fix, and one question replaces them all.

Subject
My own AI chats
Filed
16 Nov 2026
Video
12:51
Auditor
Abhishek Sinha

Findings

  1. FailedIn 97 sessions, my AI had only one job running 91% of the time. I typed "status" over 30 times.
  2. FailedSixty percent of my 633 prompts were ten words or fewer, with no why and no definition of done.
  3. FailedResearchers timed 16 experienced developers: tasks took 19% longer with AI, yet they believed AI had made them 20% faster.
  4. HeldThe engineer behind Claude Code runs five numbered sessions and waits for a ping instead of watching.
  5. Run thisPaste the audit prompt from the description into your AI, get your real score, and fix the top habit.
Fig. 1. The full case on video. Jump to a chapter:
  1. Using AI all day and still behind: I audited a week of my own AI chats
  2. The five bad bosses of AI: Hoverer, Shouter, Do-the-Needful, Quick-Call, Checker
  3. Boss 1, the Hoverer: watching ChatGPT and Claude work
  4. METR study: developers felt 20% faster, took 19% longer, then the rerun broke
  5. Boss 2, the Shouter: Toyota's five whys for repeated AI mistakes
  6. Boss 3, the Do-the-Needful: commander's intent for AI prompts
  7. Boss 4, the Quick-Call: Jeff Bezos's memo rule for AI decisions
  8. Boss 5, the Checker: Google's error budget and materiality
  9. Your bad boss score and the 100x agents question

What happened

You use AI all day, and somehow you’re still behind. So I audited myself: one week, every AI session on my computer. The count came to 97 sessions and 633 prompts.

Two numbers stood out. Of all the time my AI was working, 91% of it had one job running. One. And I typed the word “status” more than 30 times.

Here is what changed. AI stopped being a chat. Now it does the whole task while you wait, so the day you gave it one, you got promoted. You became a manager, and nobody trained you. I’m a data auditor, and my job is judging work I didn’t do myself without redoing it. That is exactly the job AI just handed you.

Every office has five bad bosses: the Hoverer, the Shouter, the Do-the-Needful, the Quick-Call and the Checker. With AI, most of us are all five. Keep a score out of five, one point for every boss you recognise in yourself.

Where the control failed

The Hoverer. You press Enter, the little dots start moving, and you watch them. In my own log, my AI worked for around 26 hours that week, mostly in 88 long jobs. After nearly half of them, I was back within five minutes, checking.

In 2025, researchers timed 16 experienced developers on real tasks from their own projects, some with AI and some without. With AI, the tasks took 19% longer. Afterwards the developers believed AI had made them 20% faster. More of their time had gone on prompting, reviewing, waiting and sitting idle.

When the same researchers tried to rerun it this year, they couldn’t: developers refused to work without AI, and between 30% and 50% admitted they had held back tasks they didn’t want to do the slow way. The researchers’ own guess is that people are probably faster now and just can’t be measured. My takeaway is not that AI makes you slow. It’s that you can’t feel your own speed, and a feeling is not evidence.

The Shouter. “No, not like that. Do it again.” It works for today. Tomorrow it’s a new chat and the same mistake. Taiichi Ohno at Toyota asked “why?” five times about a blown fuse: overloaded machine, dry bearing, weak oil pump, worn shaft, and finally metal scrap in the pump because nobody had fitted a filter. The answer wasn’t a fuse. It was a missing filter.

The Do-the-Needful. One line, no context. Of my 633 prompts, 60% were ten words or fewer. Some were just “Done, publish”, but not all.

The Quick-Call. With AI, this is the chat: twenty messages later you’ve decided something big and can’t say why.

The Checker. He rereads every line, so everything waits on his desk. With AI, you reread everything as if 100% were the target.

The evidence an auditor should ask for

Each boss has an older fix, and each fix leaves evidence you can inspect.

For the Hoverer, the fix is to send the task, turn on the notification or decide when you’ll check back, and start something else. The engineer who built Claude Code runs five AI sessions at once, numbers them one to five, and doesn’t watch any of them. His computer pings him when one needs him. Watching feels like control. It’s just waiting with your eyes open.

For the Shouter, ask why instead of repeating the correction. It’s usually one of three things: an instruction was missing, some context was missing, or a check was missing. Then write the fix where the AI will see it next time: its memory, its custom instructions, a project file. In audit terms, fixing one wrong invoice is a correction. Fixing the gap that lets it through is remediation. Mine was this: every time my AI opened YouTube to research something, the video started playing in my earphones. One line in the instructions fixed it: “Mute the player first.”

For the Do-the-Needful, borrow commander’s intent. A Prussian general, Helmuth von Moltke, wrote in 1871 that no plan stays certain past the first meeting with the enemy. So you hand down the aim, not every step. The US Army says it should run three to five sentences, short enough for a soldier to remember when the plan breaks. Your AI is that soldier: intent tells it how to guess.

For the Quick-Call, ask for a memo, not an opinion. In June 2004, at 6:02 in the evening, Jeff Bezos emailed his top team: no more PowerPoint presentations. Someone writes a memo, in full sentences, and the meeting starts with everyone reading it in silence, because writing forces better thinking. The format was never the point: the AI has to organise its thinking before you react to it.

For the Checker, Google’s engineers wrote that 100% is the wrong target for basically everything, with two exceptions: pacemakers and the brakes in your car. They pick a number, say 99.9%, and the gap is the error budget, which they spend on moving faster. The auditor’s word is materiality: an error is material if it would change the decision of the person relying on it. Set the budget by the task. A first draft has plenty of room. A client message or a payment needs close to zero, with a human signing.

The test to run this week

Four lines before any big task:

  1. Why this matters
  2. What done looks like
  3. What it must not do
  4. What you don’t care about

The last line stops the AI polishing things you never asked for. So not “Make a report on Q2 sales”, but “My manager decides on Friday whether to hire two more people. Done is one page with three numbers. Don’t change the source data. Don’t worry about design.”

Four lines at the top of every decision brief:

  1. The decision I have to make
  2. Your recommendation
  3. The top three risks
  4. What you need from me

Then add one question to every brief: what would make this recommendation wrong?

For checking, don’t be the tester, build the test. Have the AI check its own work before it says done. Run a second check every time, such as another AI, another checklist, or a total that has to add up. Then you look at the exceptions, the way audit samples after testing controls.

Finally, get your real score, not your guess. The prompt I used is in the video description. Paste it into your AI and it reads your own chats back to you. My score was two. I don’t shout and I don’t recheck everything, but I hover, and those 30-plus status messages are the habit I’m fixing.

The one question that replaces all five bad bosses comes from a developer named Kun Chen: “If I had 100 times more agents doing 100 times more work, how would I do this?” You can’t hover over 100 screens, shout at 100 mistakes, brief 100 agents in one line, take 100 quick calls or check 100 reports. At 100, there is one job left: build the system. That was the job at one.

Sources

  1. METR (Jul 2025): early-2025 AI and experienced open-source developer study
  2. METR paper: where the time went (prompting, reviewing, waiting, idle)
  3. METR (24 Feb 2026): the uplift update, and why the rerun gave an unreliable signal
  4. InfoQ: how the creator of Claude Code works
  5. Lean Enterprise Institute: the five whys
  6. CNBC: what Jeff Bezos learned from requiring six-page memos at Amazon
  7. Google SRE book: introduction (why 100% is the wrong target)