Opus 5.5 Changed How I Use AI Agents, and My Logs Back It Up

Three weeks ago Anthropic released Claude Opus 5.5, and since then I've stopped doing something I'd done nearly every day since January. I used to dictate what I wanted into ChatGPT and have it write the prompt for my coding agent. Now I just talk to the agent. That sounds like a small change. For me it isn't, because that loop was the whole shape of how I built EF-Map for nine months.

I could just tell you that and leave it there. But I've learned not to trust how a new model feels in its first few weeks, because every release feels like a step change for a while. So before writing this I had the agent go through everything I've got logged and check whether the change is real or whether it's just the honeymoon. There are three sources. WhisperTyping, the tool I dictate nearly everything with, keeps every recording, 19,212 of them since August 2025, and each one stores a screen capture of the window I was talking into. Claude Code keeps the last 30 days of session transcripts, which happens to cover the whole of Opus 5.5 so far. And the git history, where every commit Claude makes carries a co-author line naming the model.

The loop I've used since January

I've been doing agentic coding since Gemini 2, and every model family has needed a slightly different way of working. Through 2025 I talked to Copilot in VS Code directly, in small steps, and I was the memory. In January I found a loop that worked better and I stuck with it. I'd talk through what I wanted with ChatGPT and it would write a prompt. I'd paste that into the agent in VS Code, the agent would do the work and write a closeout, and ChatGPT would read the closeout and the branch through its GitHub access and check what had actually been done. Then I'd smoke test it, and round again. I wrote that up in Audit, Chunk, Implement back in June.

The window captures show how completely that took over. From September to December 2025, 92% of everything I dictated went straight into VS Code. From January to June 2026 it was 11%, with 84% going into ChatGPT. Over those six months I asked for a prompt, in phrases like "give me the prompt" or "write a prompt for the agent", 1,332 times. That's about once every 900 words I said. In my dictation post I tried to estimate this from whether a recording mentioned the word prompt. The window capture is a much better signal, and it shows the loop was even more dominant than I thought.

Opus 4.6 landed in February in the middle of all that, and it was a big jump. I said at the time "we're using Opus 4.6, brand new model", and gave the agent more room than I'd given anything before. I still think that if agentic coding had stopped developing at Opus 4.6, it would have been enough for someone like me building hobby tools. Harnesses and memory would have kept improving around it and the results would have kept getting better with the same model. The labs didn't stop though, and I kept the loop through all of it.

Prompting a better model with a worse one

Fable 5 arrived in June and that's when the loop started to feel odd. Fable sits above Opus, and I was having a cheaper model write its instructions. So what I said to ChatGPT changed. Instead of "tell it how to do this" it became "give it the goal, don't constrain it, let it figure it out". For a few weeks I talked to Fable directly more often than not, and between 1 July and 15 August more than half my words, 56%, went straight into VS Code. That's the stretch I wrote about in the unattended overnight run.

The problem with Fable was the allowance. On 16 August I said that if I hadn't had Grok for the previous two days my Fable "would have been maxed yesterday". A habit had also crept in that I didn't think much about at the time. I'd found with Opus 4.6 that it would do the coding part fine but the user-facing side would come out pretty poor, and I could steer it a bit but not all the way. The Grok agents later on were the same. On 14 September I described that flow as "I get them to build and for the most part they will build what I need but they're not very good at UI and arranging things so I follow up with a Fable pass". I had Fable rationed for exactly that job.

The Grok and Muse Spark weeks

So in mid-August I went looking for more inference. I took out SuperGrok Heavy and ran Grok 4.6 in its own command line tool and through Cursor, and on 15 September I added Muse Spark. Both did real work. A fair chunk of EO-Map from that period was built by Grok, and on 23 September I said I was "very impressed" with something Muse had just finished. On 21 September I said that any of the frontier models I'd been using recently were good enough for what I'm doing, and I still think that's broadly true.

But they needed the loop. Dictation into ChatGPT went back up to 77% of my words between 16 August and 21 September, and the prompt requests came back with it. One from 15 September sums it up: "I need you to generate a prompt for Muse now and normally I would tell you it's a capable agent, don't tell it exactly what to do". Then the closeout went back to ChatGPT, ChatGPT found what had been missed, and I asked for a prompt to fix it. External adversarial review, basically, with me carrying text between two windows.

22 September

At 18:04 that evening I dictated "So this is interesting. Opus 5.5 has just dropped." The first thing I did was run it through the loop, like everything else. ChatGPT got 94% of my words that day. And for the first couple of days the output was about what I'd have expected from the models before it. I think that makes sense in hindsight. A detailed prompt from another model turns any agent into a subagent following somebody else's plan. You get the quality of the plan, not the model.

Over the next few days I just started talking to it instead, without really deciding to. On the 24th, 34% of my words went straight into VS Code. On the 26th it was 84%. From the 28th it has been 90% or more on most days.

From 28 September to 12 October it's 87%. The last time a whole month came out that high was December, when I was dictating short instructions at Copilot and doing the remembering myself. The asks went too. In those fifteen days I asked for a prompt four times, and all four were me asking Claude to write instructions for another Claude session.

Each stretch side by side

Rates are per 10,000 dictated words, from literal phrase matches over the raw transcripts, so they're rough. 13 October is left out because it's the dictation for this article, which talks about prompts, limits and Fable a lot.

What a prompt looks like now

I'd half expected my prompts to have got shorter, and on paper they have. The Claude Code transcripts let me compare the first week, when I was still pasting ChatGPT's prompts in, with everything after. In the first week 26% of my prompts were laid out like a written spec, with headings or a run of bullet points, which is a fair sign ChatGPT wrote them because I don't dictate headings. The average prompt was 710 words and the longest 10% started at 2,175. Since 28 September only 2% look like that. The average is 308 words and the longest 10% start at 615.

That's not because I'm saying less. It's because nothing is expanding what I say into a spec any more. What goes in now is just me talking. Sometimes that's two sentences. For a bigger feature it's a few minutes, and now and then ten minutes or more, laying out what I want, what I expect it to look like when it's done and the odd architectural choice I care about, but not how to build it. Since the 28th, 25 of the messages I dictated into VS Code ran past five minutes of talking and three past ten, the longest about twenty. From September to December 2025, when I was also talking to an agent directly, it was 50 out of nearly 5,000 and none reached ten. The one I dictated for this article was 1,412 words, about twelve and a half minutes.

I talk less overall, which I didn't expect. Between 16 August and 21 September I dictated about 9,200 words on a day I dictated anything. Since the 28th it's about 6,200. But so much more of it goes to the agent that it now hears about 5,400 words a day from me, against 1,300 in September, roughly four times as much. What's gone is all the back and forth with ChatGPT about the prompt and then about the closeout.

The first week against the fortnight after

From the Claude Code session transcripts, leaving out the session that wrote this. A turn is the time from one of my messages to the agent's last action before my next one.

Steering it while it works

With the old loop I'd never interrupt. I'd paste the prompt and wait for the closeout. Now I'll send a long dictation, and a minute later remember something I forgot to say, or a bit of EVE Frontier knowledge it's not going to have, or that I don't want something done a certain way. So I just dictate another message and send it while it's working. "Oh, also generate me a thumbnail for YouTube as well as part of the closeout", during a video job on 12 October. "While you're doing fixes and deploying to production before recording the video, I notice that the hover card..." during the Log Parser work this morning.

There are 52 messages like that in the transcripts since 22 September, sent while the agent was in the middle of a task. It picks them up, folds them into what it's doing and carries on, without losing the original job. I'm fairly sure part of that is the harness rather than the model, because Claude Code holds the message and hands it over between steps. But the model still has to do something sensible with a half-finished plan and a new requirement, and it does.

Telling it whether I'm around

The other thing I've started doing is ending a prompt by saying whether I'll be at the computer. Something like "I will be at the computer so if you need a judgment call from a human on anything or direction then please just ask", or that I'm going to bed and it's an overnight job. I thought I said it every time. The logs say 21 of 76 longer opening briefs since 28 September do, and the latest week has more of them than the one before, so it's a habit I'm building rather than one I've got.

When I am around and it isn't sure, it stops and asks. It leads with its recommendation, gives a couple of options and leaves a box I can dictate into if none of them fit. It has done that 34 times since 22 September, 86 questions in all, and they tend to be the right questions. One was whether CivilizationControl turrets should keep firing at a tribemate who's on grid attacking someone, because the game only tells the turret that the pilot is attacking, not who. Another asked before adding a sovereignty wash to EO-Map, because it would change the look of something I'd already accepted. Those are exactly the calls I'd want to make myself.

The allowance

This is the part that surprised me most. I haven't hit a five-hour or weekly limit since Opus 5.5 came out, and I'm not doing any less. The transcripts agree, there isn't a single usage-limit message in any of the 131 Claude Code sessions since 22 September. In my dictation, talk about limits ran at 1.30 mentions per 10,000 words in the Grok and Muse weeks. Since the 28th it's 0.11, and the one mention was about a different subscription. On 27 September I said I was "nowhere near any session five hour limits and nowhere near a weekly limit".

Two caveats. On the first evening I did get to 88% of a five-hour window, running it with the million-token context on high, still through the old loop. And Anthropic raised five-hour limits on its Pro and Max plans the same day Opus 5.5 was released, so I can't separate the model being cheaper to run from the limits just being bigger. Probably both. Anthropic say it costs 40% less to run than Opus 5 on typical work. Either way the effect for me is the same. I'm not rationing anything, which I've been doing in one form or another for most of the year.

Fable

I've got Fable available and I haven't touched it. Apart from 1,404 Opus 5 responses earlier on the 22nd, before Opus 5.5 was out, every model response in those transcripts is Opus 5.5. The last commit across my repos with a Fable co-author line is 4 September, and since 30 September every commit in EF-Map, EO-Map, CivilizationControl and RootFit names Opus 5.5. Fable went from 10.2 mentions per 10,000 words in the Grok and Muse weeks to one in the fortnight since the 28th.

The Fable polish pass has gone with it. The hover card redo and the new link styles both shipped this month built and finished by Opus 5.5, with nothing else tidying up after it. That's the thing I'd previously rationed Fable for.

I don't know if Opus 5.5 is a cut-down Fable or something else entirely. Anthropic don't say how the two relate. What they do say is that it performs at the level of Fable 5.1 on most work. For what I do, which is hobby tools, visuals on a map and a VPS to look after, nothing that stretches a model the way serious software engineering would, it feels better than Fable to me. I wouldn't generalise that past the kind of work I do. The other thing I find hard to put into words is that I just seem to gel with it. I don't sit and have conversations with it. What I mean is that whatever I describe, the output it gives back always seems to land.

What the logs can't tell me

Three weeks isn't long, and I've been fooled by new-model weeks before. The window classification is a text match on a screen capture, so a few recordings will be in the wrong bucket. Claude Code only keeps 30 days, so I can't put an Opus 5.5 session next to a Fable session in the same harness, and the Grok tool kept its edits and summaries but not my prompts. Cycle 7 started on 29 September, so the work itself changed in the same fortnight. And Claude Code changes most weeks, so some of what I'm crediting to the model will be the harness. What I can say is that the thing I felt shows up in three separate logs, and it all turns on the evening the model came out.

So yeah. For nine months the thing between me and the code was another model. Now it's just me talking, for ten minutes sometimes, and the agent getting on with it. I've no idea how long this lasts, because every time I think I've found the workflow the models move again. If you're still running a two-model loop out of habit like I was, it might be worth trying one straight dictation and seeing what comes back.

Related Posts

Eleven Months of Talking to AI Agents is the first time I went through the dictation archive, and where the prompt-brokering habit first showed up in the numbers.

Audit, Chunk, Implement: The AI Workflow Behind EF-Map is the loop this post is about, written up while I was still in it.

I went to bed and the agents kept building is the Fable 5 overnight run, and the closest I'd come to working this way before.

A Vibe Coder Stumbles Into Software Engineering covers the overnight runs, tests and trade-offs that sit underneath all of this.