When Claude's Second Draft Isn't as Good as the First

Getting one good piece of work out of Claude isn't hard. You paste in the notes, describe what you want, adjust the wording a couple of times, and something you'd happily send comes back. That tells you Claude can do the job.

It doesn't tell you it'll do it as well next time. Six weeks later, with a different client and messier notes, the same instruction comes back a little off: a section missing, the wrong tone for this client, or a number that was never in your notes. Nothing broke. The first result just depended on more than the instruction.

If your name goes on what gets sent, that second run is what matters. Here's why it's harder, and a small experiment that shows you where your instructions are relying on you.

Why the second time is harder

The first time, the person who wrote the instruction was right there.

You knew what "the usual format" meant. You noticed a missing section and asked for it without thinking about it. You knew this client likes short updates, that last month's was too long, and that your contact reads everything on her phone. None of that was in the instruction. It was in your head while you watched the draft come back.

The second time, you're busier, the notes are different, and you don't remember any of that. Everything you supplied without noticing the first time is missing, and the draft slips in exactly the places nobody wrote down.

So the good first result was partly yours. It measured Claude plus you paying close attention. What you can rely on is whatever still works when you're not paying close attention.

If you'd like something concrete to try this on, the free Friday Update Brief includes the instruction, a sample week of raw notes and the update it should produce. The third file is the useful one, because it gives you something to compare against instead of judging by feel. It costs an email address, and the rest of this article is easier with it open.

What holds up the second time

The gap between a good first result and one you can rely on didn't close as the models got better, because it isn't about the model. It comes down to three things, and all three are about what's written down.

What "finished" means is written down. Not "make it professional," but what the finished thing contains, in what order, at what length and for whom. The test is whether someone who wasn't there could read it and judge the draft the same way you would. If only you can tell whether it's right, that's your taste, and your taste isn't in the room next time.

The instruction says what it needs. An instruction that quietly expects a figure will make one up when it's missing, and the made-up number reads exactly like the real ones. A good instruction lists what it needs and says what to do when something's missing: leave a visible marker like [NEEDED: Q3 figure] rather than fill the gap. You'll spot a marked gap in a minute. An invented number can reach the client.

It's been run more than once before you trusted it. This is the step people skip, and it's the one that finds the other two.

The experiment: run it three times

Take an instruction that gave you something good and run it again on two other weeks of real material. Use genuinely different weeks, not variations on the same one, and ideally include a week that went badly, because a quiet week hides the problems a chaotic week shows up.

Then put the three drafts side by side and mark only what changed between them.

Each change points to something the instruction didn't say:

• A section that's there in two drafts and missing from the third means the instruction never said it was required.
• Drafts ranging from 180 words to 600 mean no length was set, and "concise" isn't a length.
• A different opening each time means the shape was in your head, not in the instruction.
• A date written three different ways means nobody said how to write dates. That looks cosmetic until a client's assistant tries to file them.
• A figure that shows up in one draft and not the others means the instruction was unclear about whether it was needed. And if a figure appears that wasn't in any of your notes, fix that before anything else.

You don't need any tools for this, and you don't have to decide whether any one draft was good. You're only looking for differences between drafts made from similar material, and each difference is something missing from the instruction.

Do at least three runs. With two, you can't tell whether a difference comes from the instruction or from that particular week. A third shows you whether it happens again. It's the smallest number that tells you something, which is why it's worth the extra twenty minutes.

What it looks like on real material

Sam Rivera, from our worked example, runs funnel diagnostics for B2B SaaS companies. The weekly update for Acme Widgets is the instruction Sam runs most often, so it's the one worth testing first.

Three weeks of real notes go through the same instruction. Week one was calm. In week two a data pull was late and a session got moved. Week three was the week before the report was due, and the notes are three bullets and a screenshot.

Side by side, four things change. The blockers section appears in weeks two and three but not week one, because the instruction says to mention blockers rather than to always include the section, and a calm week has none. The updates run 190, 410 and 240 words, because no length is set, and Priya, the client contact, reads on her phone. Week three opens with what happened, while the other two open with the number, which is what Priya forwards to her CEO. And week two gives a delivery date as "end of next week" where the others give an actual date.

None of that is Claude being unreliable. Each one is a sentence missing from the instruction: the blockers section stays even when it's empty, the length is a number of words, the number goes first, and dates are written one way. Four sentences, found in twenty minutes, and the calm week on its own would have shown none of them.

That's why it's worth including a bad week on purpose. Week one gave the best-looking draft and told Sam the least.

Two traps when you read the results

Both of these feel like progress, which is why they're worth naming.

Fixing the draft instead of the instruction. A section is missing, so you add it by hand and call the run a success. Next month the same section will be missing again, and the fix you made is gone. Every correction you make by hand is a sentence that belongs in the instruction, and the time to write it in is now, while you still remember why.

Loosening the instruction until everything passes. When drafts vary, it's tempting to relax the requirement: "roughly one page" instead of a word limit, "a clear next step" instead of a named action and a date. Then every draft passes and none of them get better.

Go the other way. Where the drafts differed, be more specific. A word limit counted on the finished draft works better than "keep it short," because there's a clear answer to whether a draft meets it. A date written one exact way works better than "include a date," because there's one right answer instead of six reasonable ones.

What it costs, and what you get

Running an instruction three times and reading the differences takes under an hour. Rewriting it based on what you found takes about another hour.

What you get is something you can rely on. An instruction that only works when you're paying close attention gets worse whenever you're tired or rushed or haven't used it for a while. One that's been tested this way gives you a similar draft on a bad week, from messier notes, without you standing over it. You can hand it to someone else, or to yourself in six weeks, when you've forgotten everything you didn't write down.

It's also worth doing before you add more AI to your practice. Another instruction built on one you can't rely on just adds more variation. One instruction you trust is worth more than five that only work on a good day.

Where the written-down parts live

All three things above come down to the same idea: decisions made once and written somewhere Claude reads, instead of made again under pressure every time.

That's what the Claude Workspace Kit is: a folder holding four written standards for what finished client work contains, six briefs that say what inputs they need, and context files for your practice (who you serve, what you sell, your price floor and how you write). It comes filled in for a fictional practice, including a client folder with a week of messy notes, so you can see how it works before you set anything up. If you'd rather stay in chat, the Claude Project Kit does the same job there. The Project Kit sets Claude up to know your practice in chat; the Workspace Kit sets Claude up to do the work in your files.

Neither replaces the experiment. Standards someone else wrote still have to meet your practice, your clients and how long they like things, and the way to find out where they don't is to run them three times and look at what changed.

Pick the instruction you use most often: the weekly update, the meeting debrief, or the proposal. Run it on this week's notes and on two earlier weeks, one of them a week that went badly, and look only at what changed. Each change is something to write into the instruction this week.