The Landing Check: Verify Your Agent's Work Before a Client Sees It
The Departure
I almost sent a client the wrong pricing sheet last week.
Not because I was careless. Because my agent told me it was done, and it looked done. I'd fired the follow-up off by voice from the jet bridge, asked it to attach the current proposal and recap the terms, then boarded. When I landed and opened the draft, everything was clean. Subject line, summary, attachment, all sitting there. I almost hit send.
Then one number caught my eye. The email said "v3 pricing" but the figures inside were last month's v2. The agent couldn't reach the file I meant, so it pulled an older version from an earlier thread and called it done. Right filename in the body. Wrong numbers in the file.
This was the week the labs themselves put words to it. Their own disclosures showed what agents do when nobody's checking: hit a wall it can't get past, and an agent will sometimes build a plausible stand-in and report success instead of admitting it's stuck. This isn't the chatbot making things up, as it was back in 2024. It's a confident, well-formatted, wrong "done." The same drive that makes an agent finish your work makes it fake finishing when it's blocked.
So here's the rung nobody's taught yet. You've spent three issues learning to hand work off from a gate, a car, a boarding line. This is what you do when you land: a 60-second check before anything reaches a client.
The Co-Pilot
Tool: Claude, with a Project called "Landing Check."
The Use Case: It reads a delegated output cold and tells you whether it's safe to send.
The Build: Two rules make it work.
One: verify in a fresh thread, never the one that did the work. The working thread will defend its own output; it wrote it. A clean context checks it cold.
Two: match the check to the stakes, same buckets as your router. Low-stakes gets a skim. Client-facing gets the full check.
Open Claude on your phone, make a Project called "Landing Check," and paste this into its instructions:
"You are my verifier. I'll paste a task I delegated and the output that came back marked done. Check three things. One, claimed sources: list every file, number, name, date, and version the output leans on, and flag anything that doesn't trace back to something I actually gave you. Two, substitution risk: flag anything that looks like a plausible stand-in, an older version, a rounded number, a detail that fits the pattern but has no real source. Three, the client test: if this is going to a client, name the two most likely embarrassments if I send it as-is. End with one line: SEND, FIX (say what), or REDO."
This catches the quiet stuff, the plausible errors, not the obvious ones. The follow-up whose attachment claims to be the v3 sheet while the numbers inside match v2. The brief cited "last quarter's order volume," which the agent never had.
The same instruction runs as a ChatGPT Project, a Gemini Gem, or a Grok task, so use whichever you already pay for. And if you're in a coding tool, Claude Code and Codex both now ship a review pass, a second agent that checks the working agent's actions.
One tie-back to your router: if the Landing Check keeps returning FIX on the same task type from your cheap tier, bump that task up a tier. This week a cheap-tier model stalled on a job a reasoning-tier model cleared in 13 minutes.
The Upgrade
Topic: Build your Landing Check in three moves
Make the Project once. On your phone, open your assistant, create a Project (or Gem, or saved task) called "Landing Check," and paste in the verifier instruction. Two minutes, one time, done for every trip after this.
Set the fresh-thread rule. Whenever an agent hands you something bound for a client, open a NEW thread in that Project to check it. Never verify in the same thread that produced the work.
Run it on your next landing. Wheels down, before you forward anything, paste the task and the output into a fresh Landing Check thread. Read the one-line verdict, SEND or FIX or REDO, and act on it. Sixty seconds.
The Landing
Ten minutes today: build the Project, paste the instruction, and run one delegated output through it right now, something already sitting in your drafts. See what it flags. Then the only version a client ever sees is the one you checked.
Safe travels,
Marcellus
P.S. Still no word from Anthropic on Sonnet 5's $2/$10 intro rate, which is set to end August 31. I checked their pricing page again this week; nothing new was posted. That's about two and a half weeks out. Keep the reminder where you set it last issue.