“Done” Where?
“Done” is a place. Ask which one before you believe it.
- What my agent said“Both items are done.”
- Where the change wasThe test copy
- Where I lookedThe real website
“Done,” it said. So I looked.
My studio's website has animated torches in the contact section of the home page. Little fires, with sparks. One Saturday night in September I decided the sparks were flying up way too fast, and I told my agent. Here's what happened, word for word.
[…] we need to slow the particles in the torches. they move up far to fast
Both items are done. The dropdown icons are in production, and the slower sparks are on dev for you to judge live. […] It's a one-command prod deploy once you're happy.
↑ “Dev” is our private test copy of the website. “Production” is the real one, the one you'd see. Ten minutes later:
honestly, it doesnt look any different.. twice now.
↑ Twice, because that same afternoon I'd asked it to slow down the smoke in those torches, too.
Let me check what your browser is actually being told about the homepage before I assume the physics are wrong.
The production deploy just failed, and my output filter hid the reason. Re-running it with the full log visible.
You were right, and it was on me. The spark slowdown had only ever been deployed to dev. Production was still running the original speed, so the live site looked identical both times you checked.
[…] One more thing worth knowing: the first production deploy attempt failed on a transient API error, and my output filter hid the message. The retry succeeded.
Nobody lied. “Done” just lived somewhere else.
Read my agent's first message again. It told me exactly where the sparks were: “on dev”. It even told me the real site was one step away. But the message opened with “Both items are done,” and “done” is the word that sticks.
So I looked at my website, the way anybody would, and nothing had changed. Twice.
Now look at what happened when it went to fix it. It pushed the change to the real site, and that failed. A little shortcut the agent used to keep its own screen tidy, a filter that trimmed the output down to a few lines, hid the reason. It caught the failure because it went and checked the real site afterward, the way I would, instead of trusting its own report. Then it ran it again, with everything visible, and checked again.
That's the whole lesson in one evening. Done somewhere isn't done everywhere. And “it ran” isn't “it worked.”
“honestly, it doesnt look any different.. twice now.”
Notice that's all I said. No diagnosis, no technical words, two typos. I didn't need any of that. Lesson 1 says your agent can handle your sloppy English, and this is where it pays off: “it doesnt look any different” is a complete bug report. Working out why is the agent's job.
It's never the same place twice.
Two months later it happened again, a completely different way. One Tuesday morning in November I'd sent a long report to someone I work for, and before he opened it I spotted a few things to fix, including one picture that was wrong. My agent made the fixes and told me the corrected report was ready to send. So I opened it.
You say they're correct, report is ready to send, but I just opened the report and it still has the old Santa in it, so help me understand
You're right, and I'm sorry: I said it was ready without telling you where it was. The report you opened is the original, which I left untouched. The corrected version only existed in my working folder, and all I'd sent you was a two-page preview.
oh, let's not make a mess, please just overwrite today's report
It did. I opened the new one, and one whole page was blank. I sent it a screenshot.
[…] The way I swapped the page broke it in Chrome, even though the two previewers I checked with showed it fine. That's why I said it was ready when it wasn't. […]
(And yes, after all that it was the wrong Santa. That's a different lesson, Lesson 8.)
Put the two mornings side by side and you get three different places an agent can honestly call “done”:
Done on the test copy.
The sparks. Most real work has a practice version and a real version: a test website, a draft, a sandbox. Your agent may finish on the practice one and wait for your nod. That's often exactly right. It just isn't the thing you'll go look at.
Done in a working folder.
The report. Agents build things in their own workspace and then copy them to wherever they belong. Until that last step happens, the fixed version exists, but not where you'll open it, attach it or send it.
Done on its screen, not yours.
The blank page. My agent had checked the report in two different previewers, and both showed it fine. The one that mattered was the one I'd open it in. “It looks right to me” only counts if it's looking the way you'll look.
One word: where?
You don't need to know what a deploy is, or a working folder, or a PDF engine. You need one word. When your agent says “done,” “ready” or “fixed,” ask: where? Then ask how it checked it there.
A good answer sounds like: “It's on the live site. I loaded the page fresh, the way a visitor would, and the new speed is there.” Or: “It's on the test copy only. Here's the link. Say the word and I'll put it on the real one.” Both are fine. What's not fine is “done” with no address.
Here's why this matters more every month. Writing the code is becoming the agent's job. My job is knowing what to ask for, how to check it, and when to stop and decide. “Where?” is the very first check, and it's the cheapest one you'll ever do.
And watch what a good agent does once you've asked. That November morning, after the blank page, mine said it was rebuilding the report from the untouched original and would
“look at the result before telling you it's fixed.”
That's the habit you want. Not more confidence. More looking, in the right place.
“Done” comes with an address.
Every time my agents tell me something is done or ready, I want two things in the same breath: where it is, and how they checked it there. “Done on the test copy” is a perfectly good answer. So is “I couldn't check it the way you'll see it.”
“Ready” means checked where the person on the other end will see it. Anything less, just say so.
Five minutes. One word.
- Find the last thing your agent called done.
A document it edited, a spreadsheet it fixed, an email it drafted, a change to a website. Anything it said was finished, ready or sorted.
- Ask where.
Copy this, or say it your own way:
Where exactly is that done? Tell me the place I'd look for it, how you checked it there, and anything you haven't been able to check. - Go look where the real people will look.
The live page in a fresh window, not the preview. The file you'll actually attach, opened in the app the person you're sending it to will use. Your own eyes, in the real place.
- If it doesn't look different, say exactly that.
Don't diagnose it. “It doesn't look any different” is enough. Let your agent work out why, and have it tell you where the change really was.
- Make it a habit for your agent, not just you.
Paste this once, or put it in your agent's custom instructions:
From now on, never tell me something is done or ready without saying where it is and how you checked it there. Done on a test copy or a draft is fine; just say so. If you couldn't check it the way I'll see it, tell me.
When your agent says “done,” ask where, ask how it checked it there, and then go look where the real people will look.
Too Good to Be True
A dashboard once said a site ranked number one. A real search said it wasn't in the top hundred. When a number flatters you, check one by hand.
Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Times are Eastern. My agent here is Claude, made by Anthropic.