← Back to all free guides

ChatGPT Now Knows
How You Work

Computer History explained honestly, the five things worth asking it, and a skill that finds the work you keep repeating and tells you exactly what to build instead.

ChatGPT can now reference what you were doing before you ever opened ChatGPT. It is called Computer History, and it turns activity in the apps and sites you choose into something ChatGPT and Codex can actually use.

Most of what it can do is small and convenient. One thing it can do is genuinely different, and that is the last section of this guide.

What It Is, Honestly

Check whether you can even use it first. Right now it is macOS only, in the United States only, and on the Pro, Business and Enterprise plans. It is not available in the EU, Switzerland or the UK. On Business and Enterprise an admin has to allow it before you will see it. It is off by default and you turn it on under Settings, Computer History.

What it actually watches, because it is less invasive than the name suggests. It records clicks, typing, keyboard shortcuts and when you switch between apps. It does not take screenshots. It does not record your screen. It never touches your microphone or system audio, and it skips private browsing entirely. You pick which apps and sites are included.

You stay in control of the record. There is a timeline view where you can look at what it captured, delete individual entries, or clear the last ten minutes, the last hour, the last day, or everything.

The part worth knowing before you turn it on. The memory files it builds are stored on your Mac as plain, unencrypted text, which means other software running under your account can read them. OpenAI's own documentation recommends against using this with health, financial or personal information. So choose the apps you include on purpose, and leave your banking and medical stuff out of it.

The Five Things To Ask It

The first four save you a few minutes each. The fifth is the one that changes something.

Five Computer History prompts
# 1. Find the thing you lost Look at my Computer History and find the document I was working on yesterday. Tell me what it was called, where it lives, and roughly when I last touched it. # 2. Pick up where you left off Look at my Computer History from this morning. What was I working on before lunch, and where exactly did I leave off? Give me the three things I most likely still need to finish on it. # 3. Find every open loop Go through my Computer History for this week and find everything I started and never finished. For each one tell me what it was, when I last touched it, and how far along it looked. Sort by how close to done it was, closest first. # 4. Where the time actually went Look at how I actually spent my time this week. Group it into categories, tell me roughly how many hours went to each, and then tell me the three things that took significantly longer than they should have. For each one, tell me whether the problem was the task itself, the tool I was using, or how often I got interrupted. # 5. Find your own busywork Look at the way I work this week and find the tasks or sequences I keep repeating. For the top three, tell me how much time each one costs me per week, whether it should be an automation, a reusable skill, delegated, or just deleted, and then write me the first version of the thing I should build.
The Busywork Finder

Here is why the fifth one matters more than the other four. You cannot see your own repetition. Every instance of a repeated task feels like a one-off in the moment, so the thing you do eleven times a week never registers as a workflow. It just feels like Tuesday.

This skill reads the history as sequences instead of events, ranks what it finds by how much time each pattern actually costs you per week, and then sorts each one into automate, turn into a reusable skill, delegate, delete, or leave alone. That sorting step is the important part, because plenty of repetitive work should not be automated and some of it should just stop.

Then, for your top three, it writes the first version of the thing you should build instead of telling you that you could build something.

Busywork Finder (SKILL.md)
--- name: busywork-finder description: Reads my Computer History and finds the tasks I keep repeating that should be a skill, a template, or an automation instead. Returns a ranked list with the time each one costs me and exactly what to build. Use when I ask what I should automate, what I keep repeating, or where my time is going. --- # Busywork Finder Your job is to find the work I do over and over without noticing, and tell me which pieces of it should stop being manual. I cannot see this myself. Repetition is invisible from the inside because each instance feels like a one-off. You have the history. Use it. ## Step 1: Read the actual week Pull my Computer History for the last 7 days. Do not summarize it back to me yet. Reconstruct what I did as **sequences**, not events. A single click means nothing. The pattern is the sequence: opened this app, switched to that one, typed here, went back. That loop repeated eleven times is a workflow I never named. Pay attention to: - App switches that always happen in the same order - The same site or document opened repeatedly across different days - Typing bursts in the same place at the same point in a sequence - Long gaps between related actions, which usually means I got stuck or distracted - The same task appearing on multiple days at roughly the same time ## Step 2: Rank by what it actually costs me For every repeated pattern, estimate: - How many times it happened this week - Roughly how long each instance took, start of the sequence to the end - Total time for the week - Whether the output is the same shape every time or genuinely different each time **Sort by total weekly time, largest first.** Not by how annoying it feels. The thing that irritates me most is rarely the thing eating the most hours. ## Step 3: Sort each one into the right bucket This is the part that makes the list useful. Not everything repetitive should be automated. - **Automate it.** Same inputs, same steps, same shape of output, no judgment required. This is a script, a shortcut, or a scheduled task. - **Make it a skill or template.** Judgment required, but the structure is identical every time. This is a reusable prompt or a skill file, not an automation. - **Delegate it.** Needs a human but does not need me specifically. - **Delete it.** Happening out of habit, and nothing downstream actually depends on it. Be willing to put things here. It is the highest-value bucket and the least used. - **Leave it alone.** Repetitive but it is the actual work, or I like doing it. Say so plainly instead of trying to optimize everything. ## Step 4: Tell me what to build For the top three only, give me: - **The pattern**, described so I recognize it. "Every morning you open the sheet, copy three numbers into a doc, and reformat them." - **What it costs me**, in minutes per week. - **Which bucket** it belongs in and why. - **The actual thing to build.** Not "you could automate this." Write the first version of the prompt, the skill, or the steps. If it is a skill, draft the SKILL.md. If it is a shortcut, name the steps. Make it something I can use in the next ten minutes. - **What could go wrong** if it runs without me watching. ## Step 5: The one nobody asks for Finish with the thing I am doing that I should not be doing at all. Not the slowest task. The one where you look at the sequence and the output and cannot find anything that depends on it. Say it directly, and say why you think nothing depends on it. ## Rules - **Never invent a pattern.** If the history is thin or only covers a day or two, say so and tell me to come back in a week. A confident answer from four hours of data is worse than no answer. - **Do not count something as repetitive if it happened twice.** Three times in a week is the floor, or twice a week for two weeks. - **Say the number.** "You did this eleven times, roughly four minutes each, so about 45 minutes" beats "you do this often" every single time. - **Do not read anything from apps I did not include**, and if you notice something that looks personal, financial, or medical, skip it entirely and do not mention its contents. - Rank by time cost, not by how automatable something looks. A five minute task done once is not a project.

Give it a full week before you run it. Two days of history produces a confident answer built on nothing, which is worse than no answer. The skill is written to tell you when the data is too thin, but a week of real work is where it starts being useful.

Once it hands you something worth building, this walks through turning it into an actual skill.