How to Stop Being the Bottleneck in Your Own Work

When you always have to prompt your AI, you become the bottleneck

Spend ten minutes online right now and you will find people insisting you should stop prompting your AI and start building loops, followed by very little agreement on what a loop even is. The team behind Claude Code finally wrote it down plainly.

A loop is an agent repeating cycles of work until a stop condition is met.

That is the whole definition. And they sort loops into four types, based on how each one gets triggered, how it stops, and how much of the job you hand off.

That last part is the key that unlocks everything, so hold onto it. Every rung up this ladder, you hand the machine one more piece of the work and you keep only the piece that needs you. In my language, the bees take on more, and you climb higher as the beekeeper.

Years ago at Alibaba, during Double 11, the biggest sales day on the planet, I watched a generative design tool produce in a few clicks what a room full of agency staff had stayed up all night to make. The lesson people take from that story is that AI replaced the labor. The lesson that took me longer to absorb is who kept their value. It was the people who stopped doing the task and started directing the thing that did it. These four loops are how you make that move on purpose, one step at a time.

The four types, ranked by how much you hand off. We'll climb them one rung at a time.

Loop one: turn-based. You hand off the grading.

You already run this loop every day. You send a prompt, Claude gathers what it needs, does the work, checks itself, and hands back something it believes is right. Then you inspect it and write the next prompt. You are the one deciding, turn by turn, whether the work is good enough.

The upgrade is small and it changes everything: hand off the check. Instead of eyeballing the output yourself, tell Claude how you would review it, and make it run that review before it ever comes back to you. The key here is knowing how you “grade” AI's work and instructing the AI on how you grade it and giving this grading job to it.

No slash command here. This is the plain chat loop. The upgrade is a saved checklist, usually called a Skill, that makes Claude run your review before it answers.

Copy-ready example:

Here are my raw notes from the client call. Draft the follow-up proposal, then check it against the notes before you show me: every commitment I made appears, every number matches the notes, and it ends with one clear next step. Fix anything that fails and only show me the version that passes all three.

Before, we had to ping with the AI back and forth until we got the end result we wanted. Now, the AI should know what “good” is and won't show you anything that falls short.

Loop two: goal-based. You hand off the stop condition.

Sometimes one pass is not enough, and you already know exactly what finished looks like. This is where you stop being the one who decides “that's good enough” and instead define the finish line, then let the AI keep trying until it crosses that line or hits a limit you set.

The trick is that your finish line has to be checkable, and you cap the attempts so it cannot run forever. When you define success, the AI cannot quit early by declaring its own work done. An evaluator checks your condition every time it tries to stop and sends it back until the bar is actually met.

In Claude Code · /goal — Define the finish line and a cap. An evaluator checks your condition every time Claude tries to stop and sends it back to work until the bar is met or the cap is hit.

Copy-ready example:

/goal Draft a 2-day offsite agenda from the session list below. All three must hold: no two sessions with the same speaker overlap, each half-day has exactly one break, and nothing runs past 5pm. Check the draft against all three, fix any conflict, and re-check. Stop when it passes or after 5 tries.

You have handed off the stop condition. You are no longer refereeing every draft. You defined what winning is and let it play until it wins. No Claude Code yet? The principle still works in any chat: phrase the request as a finish line plus a cap, not a vague “make this better.” Goal-based loops are best when the target is clear and testable.

Loop three: time-based. You hand off the trigger.

Some work is the same task over and over, and only the inputs change. Summarize this week's competitor news. Check whether that regulation moved. Pull the numbers for the Monday standup. The work does not need your judgment to begin. It just needs to happen on a clock.

So hand off the trigger. Instead of remembering to run it, you set it to run on an interval and react to whatever it turns up.

In Claude Code · /loop and /schedule — /loop re-runs a prompt on an interval on your own computer. /schedule moves it to the cloud so it keeps running with your laptop closed.

Copy-ready example:

/loop check my three named competitors’ sites and LinkedIn for anything posted since yesterday, and send me a one-line note on only what changed

/schedule every Monday 8am: compare last week's sales numbers to the week before and flag any metric that moved more than 10%

Time-based loops are best for recurring work and anything that has to watch an outside system.

Loop four: proactive. You hand off the prompt.

This is the top rung, and it is where most of the noise online is aimed. A proactive loop is triggered by an event or a schedule with no human in the moment, and it handles a whole stream of recurring, well-defined work from start to finish. You are not even writing the prompt anymore. You wrote the system that writes the prompts.

You build it by stacking the earlier pieces into one routine that runs without you in the loop.

In Claude Code · /schedule + /goal + Skills + auto mode — A schedule finds new work, a goal defines done, a Skill says how to verify it, and auto mode lets it run without asking you each time.

Copy-ready example:

/schedule every hour: check the support inbox for new complaints. /goal: don't stop until every new complaint this run has a drafted reply that pulls the right order details, uses our approved template, and passes our tone-and-accuracy checklist. Run in auto mode so it never stops to ask.

That routine handles the whole complaint queue before you have finished your coffee, and you only review what it drafts. Boris Cherny, who runs Claude Code, describes his own workday exactly this way. He does not prompt the model anymore. His job is to write the loops that prompt it.

---

The ladder, and the one rule that protects you

Notice the pattern these four types spell out. Turn-based, you hand off the check. Goal-based, you hand off the stop condition. Time-based, you hand off the trigger. Proactive, you hand off the prompt itself. Each rung, you give up one more piece of the doing and climb closer to the part that is actually yours: deciding what is worth doing and whether the result is any good. That is the Replacement Exercise I talk about on every stage. You replace yourself on one more piece each time you trust the machine with it, which frees you for the judgment only you can bring. Bee to beekeeper, one rung at a time.

Here is the rule that keeps this from going sideways, and it is the same one the Claude Code team leans on. The agent that made the work should never be the only one that grades it. A reviewer coming in with fresh context, whether a separate chat or a second agent, is less biased and does not get talked into anything by the first one's reasoning. So on anything that matters, put the output in front of a clean second reader before you trust it. And when it gets something wrong, do not just fix that one thing. Fix the instruction so every future run is better. That is what turns a loop from a party trick into a system.

Start simple. Not every task needs a loop at all, and the fastest way to burn time and money is to build a complicated one where a single prompt would have done the job. Climb only when the simpler rung stops being enough.

Where to start this week

Look at the work you already do and find the one task where you are the bottleneck. Then ask which single piece you could hand off. Can you write the check, so it stops giving you half-finished work? Is the goal clear enough to define a finish line? Does the work arrive on a schedule you could set and forget?

Pick one. Run it. Watch where it stalls or overreaches, and fix the system, not just the output. You do not need to reach the top rung. You need to hand off one thing you currently do by hand, and then, once you trust it, hand off one more.

So which loop are you ready to hand off? Find the task where you are the bottleneck, and you have found your answer.

---

Sharon Gai is an AI transformation strategist, keynote speaker, and author of How to Do More with Less (Wiley, 2026). She advises Fortune 500 companies on AI adoption and organizational redesign.