Lab notes · 4 jobs
Claude is for the document you still have to sign
Claude is a general model that behaves like a careful reader. You feed it a long file and it stays in the file longer than most tabs. That is the product: context window and tone, not plugins, not a brand CMS. I ran four jobs: a digest of a 28-page vendor MSA, a rewrite of a founder letter that had gone sharp, a pass over a 400-line Python module, and a request for forty Facebook variants I asked for on purpose. Two jobs are why you would pick this over ChatGPT. One is why you still read the redline. One is why a performance marketer should not.
Run 01
I asked it to digest a 28-page vendor MSA
What I fed itI dropped in a services agreement: auto-renewal in section 8, a cap on liability in 12, a data-processing addendum that contradicted the security exhibit. I asked for a one-page brief a GC could skim, with section numbers.
Where it shone
It kept the section numbers. It put the contradiction on the first page, not in a polite appendix. Most models summarize the heading and miss the fight in the exhibit. Claude found the fight.
Where it face-planted
It softened the cap. The draft said the vendor “limits certain damages.” The contract said consequential damages are excluded, period. Softening is how a careful model tries to be fair. Fair is not the job. I put the exclusion back in the first sentence.
The useful output was the map of the document. The dangerous output was a kinder version of a clause that is not kind. Claude will read the whole thing. You still have to refuse the rewrite of the risk.
Run 02
I asked it to rewrite a founder letter that had gone sharp
What I fed itThe original thanked no one and blamed a missed quarter on “a team that did not row.” I asked Claude to keep the facts and lose the sneer. Same length. No TED-talk ending.
Where it shone
It kept the miss. It did not invent a silver lining. The tone came down without turning into mush. That is the careful-prose seat: you can send it.
Where it face-planted
It added “we are learning.” Nobody had said that. A hedge in the last line is how Claude makes a letter safe. Safe is not always true. I cut the lesson and left the miss.
I kept the temperature. I cut the moral. Claude will save you from a letter you will regret. It will also save you from a letter that needed a spine.
Run 03
I asked it to review a 400-line module I already understood
What I fed itA payment-retry helper with a silent catch, a magic number, and a name that lied about what it returned. I asked for bugs first, style second.
Where it shone
It led with the silent catch. It did not open with import order. If you already know the file, that is the useful colleague: it reads for failure, not for lint.
Where it face-planted
It then spent a page on whether the helper should be a class. That is a taste argument. I did not ask for architecture. Past the first finding, Claude will keep talking because it is a good student.
Take the first real bug. Ignore the seminar. Claude is a better reader than it is an architect-for-hire.
Run 04
I asked it for forty Facebook variants — the wrong job
What I fed itI asked for forty headlines and forty primary texts for a retargeting ad on a $49 course. Same offer, different hooks. This is a volume job. Claude is a long-document seat. I ran it so we would have a clean miss.
Where it shone
Nothing I would buy Claude for. It can write one good line. It will not write forty that are actually different.
Where it face-planted
By variant twelve they were rearranging the same four words. By variant thirty they had grown a lecture about “authentic connection.” A media buyer needs a spreadsheet, not a sermon. I stopped at fifteen and opened a cheaper tab.
This is why Jasper and Anyword exist, and why Claude should not pretend to. If the job is a file you will sign, stay. If the job is a grid of hooks, leave.