ISSUE 18 · JULY 21, 2026 · ME+MACHINE
Today I started a prompt with the instructions "Follow the process defined in this file to draft a brief." Claude asked me a few clarifying questions, started whirring away, and then... didn't follow the instructions.
What? Why? How???
The process was clearly defined in both the prompt itself and the standing instructions Claude reads at the start of every session. And the process wasn't complicated either: validate the idea → research the topic → agree on the approach → then write.
Claude read every word, yet skipped straight to "write" four times in a single afternoon. Each time it got caught, it apologized sincerely, explained the process correctly, and then did it again. So what was happening?
AI doesn't execute instructions, it weighs them. Every reply is the AI picking the most likely next words to write, and your rules go into that pick as just one influence among many. When another influence weighs more, the rule loses. Nothing physically stops the AI from skipping a step the way a checkout page stops you without a credit card number.
The finished deliverable weighs the most. An AI is built to hand you completed work, so the strongest pull is to skip ahead to the end. Deadline pressure adds to it. So does pushback. In fact, every time I challenged Claude about a skipped step, it came back with another finished draft.
And your rules weigh less the longer the work runs. Early in a session the rules are fresh. Twenty minutes into building something, they're background noise competing with everything else the AI is weighing. A study from this spring found that models could repeat a rule back correctly in the same exchange where they were breaking it, and rule-following tended to drop as the back-and-forth stretched on.
In other words, a process is the hardest instruction you can hand an AI: a chain of rules standing between it and the finished deliverable, and every link is another chance to skip ahead. Writing the rules more clearly, or more forcefully, won't solve that. In fact, at one point, I told Claude to "make sure it never happens again," and watched it skip again in the same conversation.
The fix is structural, not editorial. For me, that meant putting a gate between every step. Instead of asking Claude to follow the process, I asked for each step's results, one at a time. A request for only step two’s output is a request it can't skip. "Show me the research you gathered, and nothing else. I'll tell you when to move on to the next step" cannot produce a premature draft, because you didn't ask for one.
Here are a few good habits to use in your own workflows:
Make each step produce something you can see. A step isn't done when the AI says it's done; it's done when you've looked at its output. You are the enforcement.
When it messes up, go back to the skipped step, don't fix the output. Editing a draft that skipped the research just polishes the wrong work. This was my expensive lesson of the afternoon: I kept reacting to patched drafts (four of them!) when I should have been sending Claude back to the step it skipped.
Fresh eyes to check, never the author. The conversation that wrote something shares the blind spots that wrote it. A study this March found that a review run in a separate session catches mistakes that same-session reviews miss. I experienced this as well when a separate chat caught problems the drafting conversation had read right past.
Tell it what to do, not what to avoid. "Open with a real number or a quote" works while "don't be generic" doesn't, because the AI can satisfy a list of don’ts a hundred bad ways (even Anthropic's own prompting guide says to phrase it this way).
Here’s a gate you can add to any multi-step prompt:
PROMPT
We're doing this in steps, with a gate between each one.
The steps, in order: [list them].
Start with step 1, and do step 1 only. End every step by showing me what it produced, then stop and wait until I say to continue. If you notice you're producing something I haven't asked for yet, stop and tell me which step you think you're on.Copy this into ChatGPT or Claude and tweak the wording to fit your process.
You don’t need to use this prompt for a quick question. A gate is for work with multiple steps, especially where a skipped step is easy to miss or has unintended consequences. If you’d rather one-shot your prompt and have the AI run all the steps in one go, that works too. Just tell it to label what each step produced on the way to the final answer. That way, if it skips any steps, you’ll see the gaps.
THE 30-SECOND VERSION
An AI can know your written process, explain it back correctly, and still skip steps. Knowing the rule and following the rule are two different jobs.
A written rule is weighed as one influence among many, not enforced like a checkout page.
The pull toward the finished deliverable is usually the strongest force in the room, and challenging the AI on a skip may just get you another finished draft.
Writing the rule more clearly didn't fix it for me; the same rule broke again in the same conversation.
The fix: ask for the step, not the process. Each step ends in output you look at before the next one is allowed to start.
My brief did eventually get written, one gated step at a time. And I came away with a lesson of my own, because Claude wasn't the only one skipping steps that afternoon.
Your turn. Have you caught your AI skipping your instructions? Hit reply and tell me what went wrong.
me+machine