Last week I almost deleted my AI agent
Tuesday, 11pm. I asked AI to sort through 20 emails from clients so I could review them the next morning. It mixed up two client names, changed the contract amount three times, and I ended up spending 40 minutes double-checking everything — slower than if I'd just done it myself. That night I genuinely wondered: can this thing actually help me, or is it just a fancy toy?
Then I read how Linear's team builds theirs
Linear is a $1.25B company. Inside they have something called "Linear Agent," and their product lead Nan Yu plus engineer Jacob Shumway shared 5 rules. The one that hit me hardest: don't make AI start from scratch — plug it into workflows you're already running.
Example: if your work always starts in a WeChat group, let AI read group messages directly. Don't make a separate page where you copy-paste stuff in. My friend Lin used to do Xiaohongshu content; she had AI read her saved viral posts and spit out a v1 draft in two minutes — no doc library, no fancy setup. I'd gotten this wrong before. I kept trying to make AI "proper," and the more proper it got, the less useful it became.
What it costs to replicate today
Money: $0. Free ChatGPT or free Claude can test rule #1.
Time: 1 hour. Pick the task you repeat most today, write out the whole flow: where it starts, what tools it touches, what "done" looks like, where you need to approve.
Technical barrier: Zero code. You just need to know how to copy-paste a conversation. Linear's core trick is making AI "find info" instead of "remember everything" — e.g., have AI search keywords in your chat history, not dump the whole history into context.
First step: Open a work WeChat group right now. Pick one client message, copy-paste it into AI, and add: "Next time I send a message like this, draft a reply in this format. Wait for my OK before sending."
How to use this rule at each stage
Just starting out (no clients yet): Test on your own stuff first, never on clients. List anything you do 3+ times a day, pick one, let AI draft it, you just hit send. Failing is fine — AI messes up, you fix it in seconds, cost is zero.
Got 1-2 clients: Use AI for "organizing" work, not for talking to clients directly. Summarizing meeting notes, categorizing client needs, drafting first versions of quotes — these three I've tested, lowest error rate, and the time saved is enough to take on one more client.
Scaling up (team of 5+): At this stage you don't need more AI — you need to write your working AI flows into SOPs. Our mistake was letting every new hire figure out AI on their own. I fixed it by pinning our 3 daily AI templates in a shared doc, and new hires got productive on day one. If you're scaling, this step beats any new tool.
Totally fine to not try this today — it's not for everyone. But if you're drowning in repetitive work again, spend one hour on rule #1. Worst case, you sleep better tonight.