You've done this at least once: the meeting ends, you open the recording, and you type “hey, can you summarize this for me?” It works. And yet somehow, at the end of next week's meeting, you're sitting there typing the exact same sentence again.
You use AI every single day, and your evenings still disappear. Here's the thing I keep running into — if AI only ever lives in your chat window, repetitive work never actually shrinks. Below is the ladder of six levels, plus how to figure out which rung your work belongs on. (Current as of July 2026)
- The reason your workload isn't shrinking has nothing to do with how smart the model is. It's where you're using it.
- Chat → structured prompt → skill → Work/Cowork → vibe coding → automation. The whole game is finding where a given task sits on that ladder.
- Three questions decide it — how often does this repeat, is the process the same every time, and is the output format fixed?
Why AI isn't actually cutting your workload
Most of us treat AI like a smarter search bar or a first-draft generator. Ask, get an answer, copy the useful bits, paste. That's where it stops.
The problem shows up at the next meeting. Last time you got a clean list of decisions. This time, attendee opinions and final calls are mashed together, and every task has an owner but no due date.
The model didn't get dumber. The output keeps changing because you asked off the top of your head, differently, every time. Wobbly input, wobbly output.
So the gap isn't the AI's reasoning. It's that you never connected the AI to how you actually work. That's where this gets interesting.
Which level is your work on right now?
When people pick an AI tool, the first question is usually “which AI should I use?” Wrong order.
The real first questions are these three: what problem am I actually solving, how often does this come back around, and are the steps and judgment calls the same each time? Answer those and you're already halfway there.
Here's the whole ladder in one table. Pick one task you're doing this week and find its row.
| Level | Use it when | In meeting-notes terms |
|---|---|---|
| Plain chat | One-off or occasional work | Roughly summarize this one meeting |
| Structured prompt | Format and criteria matter | Specify role, order, and output format |
| Skill | You keep pasting the same prompt | “Apply the meeting-notes skill” — one line |
| Work / Cowork | Multiple sources need connecting | Notes + task table + follow-up email in one go |
| Vibe coding | Identical process every week | Drop in a file, get the finished doc |
| Automation | Timing or trigger is fixed | Save a recording to a folder, it handles itself |
Most people are parked in the top row or two. The further down you go, the less manual work you touch — but you don't need to ride it all the way to the bottom. More on that later.
Where chat, prompts, and skills part ways
“Summarize this meeting” and “You are a project manager. Using the attached transcript, organize it in this order: meeting purpose → key discussion → confirmed decisions → open questions → tasks by owner → due dates. Do not infer anything that isn't in the transcript” are not the same request at all.
The second one is a structured prompt. The point isn't writing more words — it's handing the model your judgment criteria in plain sight. Clear criteria, consistent output.
If a prompt is “how to ask well once,” a skill is “saving that good way of working as your team's standard.”
But here's what happens once you finally land on a great prompt: now you have to dig up that giant paragraph and paste it before every meeting. Manageable when it's just you. The moment your team shares it, everyone tweaks it slightly, and the outputs drift apart again.
That's your cue to move up to a skill. A skill isn't just a stored prompt — it bundles the process, the judgment criteria, the output format, and any reference templates into one reusable way of working. As of July 2026, Claude introduced them first, and you can register and use them in ChatGPT too.
Once a skill exists, the whole request becomes “apply the meeting-notes skill to the attached file.” For you personally, that means never rewriting the prompt. For a team, it means a new hire works the way your most experienced person does. The most common reaction I see in reviews is some version of “we should have rolled this out to the team months ago.”
Directing each step vs. handing off the whole job
Even with skills, you're still the one doing the shuffling. Upload the file, copy the answer, move it into a table, ask again for the email draft. You're running every leg of that relay yourself.
That's what ChatGPT's Work and Claude's Cowork are for. Regular chat can analyze files too, so what's the difference? It comes down to who breaks the job into steps and links them together. The table makes it obvious.
| Plain chat | Work / Cowork | |
|---|---|---|
| Who splits up the work | You | The AI |
| What you do | Direct one step at a time | State the end goal only |
| Connecting the pieces | Your job | Feeds each result into the next step |
| Best fit | One-off tasks | Multi-source, multi-part projects |
In chat, you personally walk it through “upload transcript → meeting notes → now a task table → now a follow-up email.” With Work or Cowork, you hand over the goal — “use the meeting-notes skill on this transcript, build a task table by owner, and draft follow-up emails to attendees” — and it splits the work up and delivers the finished pieces. It's much closer to an agent that sees the job through.
The more sources you need pulled into one coherent thread, the bigger the win. But if it's a once-in-a-while task, just hand it off that one time and move on.
If it's identical every week, build the tool
Here's the test for going one step further. If you've used Work or Cowork a few times and the inputs and requests are meaningfully different each round, don't build anything. Just delegate it as it comes.
But if the input files, the process, and the output format are carbon copies week after week, that's a different story. At that point you're better off freezing the work into an actual program.
This is where vibe coding comes in — using Claude Code on the Claude side or Codex on the ChatGPT side. The name sounds like developer territory, but the real point is turning something a person repeats by hand into a work tool anyone can run the same way.
You just describe it out loud: “Build me something where I drop in a meeting recording plus the attendee list and project name, and it spits out meeting notes in our company template along with a task table by owner.” The tool writes the code, creates the files, fixes its own errors, and tests whether it actually runs. No programming language required.
One caveat, though. Coding is not the mandatory next step for every task. It only makes sense for work that's genuinely repetitive and standardized. Building a program for something you do twice a year is just busywork with extra steps.
The last line of automation: the button a human has to press
Building the program doesn't mean you've automated anything. If a person still has to upload the file and hit run, that's not automation — that's a very good work tool.
It becomes automation when you attach a trigger or a schedule. A recording lands in a designated folder, the program detects the new file, generates the notes, sorts the action items, files the output in the project folder, and leaves the email sitting in drafts.
You could also set it to gather the week's inputs every Friday afternoon and produce a draft status report on its own. Nobody types the same request ever again.
There's one line you should never cross, though: don't automate anything that's hard to undo.
- Gathering sources, categorizing content, writing drafts → safe to automate
- Actually sending emails, deleting files, making payments, changing critical records → keep a human on the final check
Machine handles it up to the draft; a person presses send. Holding that line is the only way automation stays trustworthy long-term.
So what do you do today?
You don't need a grand rollout. Just think of one thing you'll end up doing again this week.
Ask three questions about it. Does it repeat often? Is the process the same every time? Is the output format fixed? If all three are yes, that task has no business sitting at the chat level. It's earned at least one rung down — the skill level.
Being good with AI doesn't stop at asking good questions. It runs all the way through saving what works, handing off whole jobs, and turning repetition into a system. Move down one rung today. That single step might be what changes next Friday night.
