Claude vs ChatGPT for Work: What a 1-Hour Guidebook Session Truly Taught Me

Quick disclaimer before anything else: this is one person’s workflow, not a lab test. I’m not claiming ChatGPT structurally can’t do what I’m about to describe — only that this is what happened to me, on this task, this week. Take it as one data point, not a verdict.

The hour I spent explaining the same thing three ways

I opened a new ChatGPT conversation and said something like “let’s start with the structure” — the same starting point I’d use for a Claude meeting-notes prompt, just with a different model. I was building a guidebook for a work process — the kind of document that needs both a clean structure and the actual step-by-step detail underneath each section.

It gave me a structure. Good one, actually — clean headers, logical flow. So I asked it to start filling in the actual content under each section. What came back was still structure: slightly longer headers, a one-line summary under each, but none of the specific steps I needed.

I corrected it in plain language — “I need the actual instructions here, not a summary of what this section covers.” It responded like it understood: acknowledged the gap, said it would fix it. The next output was structurally identical to the one before. Same shape, same missing depth, just reworded.

I tried this maybe four or five times over the hour, different phrasings each time. Every response agreed with the correction. Every result exported to a .md file looked the same as the first one.

Why it stuck (a guess, not a verdict)

I don’t know the mechanism behind this well enough to claim I’ve diagnosed it. What I can say is what it looked like from the outside: the first structural answer seemed to set a shape for the rest of the conversation, and no amount of natural-language correction moved it off that shape — only starting a completely new conversation did.

That’s the part I’d call an anchoring effect, loosely borrowing the term — an early response setting a frame that later corrections describe wanting to leave, without actually leaving it.

Where Claude holds a project together

The contrast that stood out: with Claude, when I keep a master file (current state, next steps) and an archive file (detailed reasoning, past decisions) attached at the start of each session, a brand-new chat picks up the thread without the same stuck-shape problem. I can open a fresh conversation days later, attach both docs, and it continues from where the project actually is — not from whatever shape the first message that session happened to set.

That’s not a claim that Claude is “smarter” at the task — it’s that the continuity comes from the files, not from the model holding it in memory across one long unbroken conversation. Which, incidentally, is exactly the structure behind the model-switching problem I wrote about last week — same underlying fix, different symptom.

This is really an extension of picking the right AI for the right job — except here the “job” is staying consistent across a single long project, not choosing per-task.

That’s really what this Claude vs ChatGPT for work comparison comes down to.

The Claude vs ChatGPT for Work Split I Actually Use

I stopped trying to make one model do the whole job — the same conclusion I landed on after comparing all three models head to head a few weeks back. What I actually do now, on longer projects:

• Strategy and structure with Claude — master + archive files, session-start prompt, the whole continuity system.
• Short creative bursts with ChatGPT — image generation, one-off brainstorms, anything that doesn’t need to survive past a single conversation.
• Export, then hand back — when a ChatGPT session produces something worth keeping, I export it as a .md file and feed it into the Claude project, so it becomes part of the continuity system instead of living in a conversation I’ll eventually lose track of.

Neither tool is doing the other’s job. The .md handoff is the seam — and it’s a manual one, on purpose. I’d rather control what crosses over than have it happen automatically and miss something.

This isn’t a ranking of which model wins. It’s closer to a division of labor I landed on after enough sessions like the guidebook one — and it might not match how you work at all. If you’ve got a different split that works for you, I’d genuinely like to hear it.