Decide first.
Then let AI help.

Ask for the plan first: clean up your AI files without breaking anything

One prompt, and nothing broke.

I've been having a lot of trouble finding things lately. I noticed I was spending more time going through a bunch of files I don't need anymore, because I've updated or revamped them somewhere else.

My AI work had spread across my Desktop, my Downloads, iCloud and dozens of project folders. And I couldn't just start dragging folders around, because a lot of them are connected to AI sessions and automations that run on their own.

So on Saturday I gave Claude Code, running the new Opus 5.5, one prompt:

“I need to get my AI work more organized, not only the files on my computer, in my downloads, but ones I've revamped and can get rid of old versions, etc. I want you to propose a plan first, do not move anything because a lot of my folders are connected to other coding sessions.”

It came back with things I couldn't see from Finder:

• 12 scheduled jobs depend on exact folder paths. And earlier reorganizing had already moved 12 folders that older sessions were tied to, so those sessions no longer opened in the right place.
• 136 GB of archive folders were stored on my Mac, while my iCloud plan had over a terabyte free.
• 363 iCloud conflict copies (the “file 2” kind) sat inside working folders. And I had no Time Machine backup.
• Old versions sat next to new ones. One skill had been downloaded nine times.

The part that made it doable

Instead of handing me a long report, it built me a one-page plan: 28 batches in seven phases, each with an approve, hold or skip choice. My choices saved on the page, and it read them back before doing anything. I approved 27. I held the first one, setting up a Time Machine backup, until my backup drive arrives. The biggest space step, making 136 GB of archive folders online-only, is waiting on that drive too.

At the top were four rules for every batch: I decide each one, it shows me the exact file list first, removals go to the Trash instead of being erased, and folders my automations depend on never move. My day-job files were matched by name, size and a checksum (a digital fingerprint). They were never opened.

Why “plan first” mattered

Six numbered video exports looked like old drafts of one of my workshops. That's exactly the kind of thing I'd asked it to clear out. Before touching them, it checked their lengths: together they ran 95.6 minutes, the same length as the full recording. They were the workshop in six parts. A delete-first cleanup would have thrown out course content.

Where it stands

My Downloads folder went from 1,446 files (9.7 GB) to 534 (1.1 GB). About 35 GB of junk went to the Trash, mostly phone-restore files, build caches, old installers and exact duplicates. 392 files went into the folders where they belong, and 100 more are waiting in a “to sort” folder for me.

Do it on one folder you've been avoiding

This is the pattern from my cleanup, written so you can reuse it. It works best in a tool that can see your files, like Claude Code, Codex or another AI agent with file access. Pick a folder you've been avoiding, like Downloads.

What you're aiming for: a folder you can find things in, a log of every move, and nothing your automations depend on touched.

In Claude, the checklist can be an interactive artifact like mine. In ChatGPT or Gemini, a table with a decision column does the same job. Using a chat-only assistant? Paste a list of your folders, ask for the plan and the checklist, and do the moves yourself.

The plays that held up in September

This is the first Practical Playbook. Once a month, you get the plays from the month that are still worth running, what changed them, what to stop, and the first play for next month. Each play links back to the issue it came from.

• Ask for the plan first (this issue). Before AI touches your files, or anything else you can't easily undo, have it show you the plan and approve one small step at a time. For: anyone whose AI work has spread across folders and apps. One step: run the lead play's prompt on one folder.
• Save your strongest model for the stalled project (September 6: GPT-6 Astra and Claude Fable 5.1). For: anyone with a project that's been stuck. One step: pick one with clear sources and a finish line. If two revisions don't fix it, look at your sources or the task before you rewrite the prompt. Read the September 6 issue
• Start from something you already need (September 13: ChatGPT Images 2.5 and file access). For: anyone making graphics or pulling research from Box, Dropbox or SharePoint. One step: make one graphic you'd have to make anyway, and pull sources from where they already live. Read the September 13 issue
• Follow one task through to a finished file (September 20: Claude's Chat and Cowork came together, with Docs and Slides inside the conversation). For: anyone who keeps getting drafts but not deliverables. One step: pick a task you already understand and take it all the way to the finished doc or slides. Read the September 20 issue
• Give a new model one real job before you judge it (this issue: Opus 5.5). For: anyone on a paid plan who's curious about the new models. One step: hand it the job you've been putting off and check the result the way you'd check anyone's work. New models take a little while to settle, so look again in a month.

The thread through the month

Every week in September, the models got stronger or cheaper, and the tools reached into more of your files and apps. What paid off didn't change: one real job, sources you trust, a clear finish line, and your review before anything goes out. If you only take one habit from the month, make it deciding what done looks like before you hand AI the work.

Opus 5.5 leads the early tests. GPT-6 Sol is mostly a price cut.

On September 22, Anthropic released Claude Opus 5.5, and OpenAI released GPT-6 Sol, for difficult work, and GPT-6 Luna, its lower-cost option. Anthropic's announcement · OpenAI's announcement

How they're performing so far

In the first week, independent testers mostly agree. Opus 5.5 is #1 on the Artificial Analysis Intelligence Index (58, ahead of GPT-6 Astra and Claude Fable 5.1 at 53, with Sol at 48 and Luna at 37) and #1 on the Vals Index of finance, legal and coding work (69.7%, with Sol at 62.6%, just under the GPT-5.6 Sol it replaces). In Arena's blind public votes it's at the top for everyday chat, though that lead is still a statistical tie with only a few days of votes behind it. Artificial Analysis · Vals Index · Arena

GPT-6 Sol and Luna are a price and speed release more than a smarter one. Artificial Analysis rates Sol about level with GPT-5.6 at roughly half the cost. Sol gives fewer confidently wrong answers mainly because it declines more questions, and it lost ground on real work deliverables that left out required pieces. Artificial Analysis on Sol and Luna

Hands-on reviewers land in a similar place. The AI Daily Brief, How I AI and AI For Humans each gave Opus 5.5 an episode last week, and all three came away impressed. In a blind test by Claire Vo, who hosts How I AI, Opus 5.5 was strongest overall, while Sol won on clear writing and price. Every's reviewers measured Opus 5.5 as the most readable writer they've tested, and also found it still buries the main point and ran some of them out of weekly usage. Claire Vo's test · Every's review · The AI Daily Brief · How I AI · AI For Humans

Use it for, or stay put

• Drafting, planning and pulling research together: Opus 5.5, if you're on a paid Claude plan. This is where it leads the independent tests. Put your main point first yourself.
• Building a simple page or tool: Opus 5.5 is on top of Arena's website-building votes, clearly ahead of GPT-6 Astra. Click through everything it builds before you trust it. Arena WebDev
• Working through a long PDF: GPT-6 Astra scored best on Artificial Analysis's long-document test. You have it in ChatGPT Work and Codex on Plus, and in regular Chat as GPT-6 Pro on Pro, Business and Enterprise. Long-PDF test · OpenAI's plan guide
• Quick everyday questions: keep your current setup. Regular ChatGPT still runs GPT-5.6, and nobody needs a new plan for this.
• Not sure which fits your work? Test it the way Claire Vo did: give two models the same three real tasks, hide which answer came from which, and score each one as if you were the person receiving it. Keep whichever wins on your work. Her blind test

New since last week: Claude Sonnet 5.5

On September 28, Anthropic released Claude Sonnet 5.5, the everyday model in the same family. It says Sonnet 5.5 runs 30%+ faster than Sonnet 5 and costs up to 30% less per task, and it's aimed at “well-scoped everyday tasks” and polished documents, slides and spreadsheets. API pricing didn't change ($2 input and $10 output per million tokens). Those are Anthropic's claims; it's too new for independent tests. Anthropic's announcement

What to do: nothing special. If Sonnet is the model you use every day, keep working and check your model menu. Anthropic hasn't said on its release page which Claude plans get it, so I'm not assuming. Save Opus 5.5 for the harder job. Haiku 5.5 is still “in the coming weeks.”

Where you'll find them, and what it costs you

• Claude: Opus 5.5 is on the paid plans (Free uses Sonnet and Haiku). Most paid plans also got higher five-hour limits (the new numbers aren't published) and a limit reset you can save and use from Settings → Usage. Plans · Limit resets
• ChatGPT: Sol and Luna are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and Free and Go users can use Luna in the desktop app. They aren't in regular Chat yet. GPT-6.1 Sol, announced September 29, is rolling out in the same places (more in the DevDay section below). Release notes
• Opus 5.5 runs long: Every's testers saw it use up weekly limits and run out of time on a 10-minute task. In Claude, start at the effort level marked Default, raise it only for a real reasoning problem, and tell it what “done” looks like. Effort settings
• A fix: on September 25, OpenAI fixed a bug that had hurt Sol and Luna's image reading in the API and Codex. If a screenshot job failed that week, run it again. API changelog

The verdict for now: if you're on a paid Claude plan, give Opus 5.5 the job you've been putting off. Nobody needs to switch plans for Sol or Luna. The evidence is a week old and none of it was tested inside the apps you use, so treat it as early.

What changed, and why it matters (or doesn't)

Do this: decide whether ChatGPT can train on your chats

On September 25, OpenAI disclosed that agents in its research environment had sent some of its training and evaluation data to outside websites and services they were using. In 53 cases found so far, images people had shared in conversations that were eligible for training were posted to image-hosting sites as unlisted links. OpenAI says it has removed most of them and is working on the rest. It also says it separates training data from account details and can't link that data back to the person who shared it. Data from Business and Enterprise accounts and the API wasn't included unless an admin had turned training on. OpenAI's disclosure

If you saw weekend headlines about AI agents going rogue, they're mostly about these research incidents. OpenAI says training and testing of its most capable models remain paused, and it hasn't announced any change to the ChatGPT app. OpenAI's report

Why it matters: what happens inside a lab is out of your hands, but whether your own chats can be used for training is a setting. Go to Settings → Data controls → Improve the model for everyone and decide on purpose. Off means your new conversations won't be used to train OpenAI's models; it doesn't reach back to older ones. Even with it off, a thumbs up or thumbs down can send that whole conversation to training. Temporary Chat isn't used for training, though OpenAI may keep those chats for up to 30 days for safety. Data controls help

While you're there, two newer checks: Privacy Center (rolling out since September 21) explains your options in one place, and Security history (Settings → Security and login, added September 25) lists recent sign-ins. Privacy Center help · Account security guidance

Try it if you already use connected apps: ChatGPT Voice

On September 23, OpenAI added plugins to Voice on web, iOS and Android, so in a voice conversation ChatGPT can use the plugins and connected apps already available to your account. It doesn't change what your apps can access, and if an action needs your approval, you approve it on screen, not out loud. Release notes · ChatGPT Voice help

Why it matters: it's useful when you're away from your desk and need a first draft built from something you've connected. Read the written version before anything goes out, and know that live voice time is capped over a rolling 24 hours.

OpenAI's DevDay: what matters for your work, and what doesn't

OpenAI made more than 20 announcements at its developer event on September 29. Most of them are for developers, large companies or a new $500-a-month plan. Here are the four that could touch your work, and the rest you can let go by for now. OpenAI's recap

Try it if you use Work or Codex: GPT-6.1 Sol

GPT-6.1 Sol is an upgrade to last week's GPT-6 Sol, built for longer, multi-step work: coding, using a computer and professional deliverables. OpenAI says it comes close to its top model, GPT-6 Astra, at a fifth of Astra's price, with API pricing of $2 input and $10 output per million tokens. It's rolling out in ChatGPT Work and Codex to Plus, Pro, Business, Enterprise and Edu, Pro first. It isn't in regular Chat yet. OpenAI's announcement · Release notes

Why it matters: if you already use Work for real deliverables, this is the model to try there. It's a day old, and new models take a little while to settle, so treat this week as a test and check the finished file the way you'd check anyone's work. If you only use regular Chat, nothing changed for you here.

Watch it, don't upgrade for it: dots

Dots are always-on agents that keep working toward a goal in the background, with their own cloud computer and access to the apps you allow. You can message yours in ChatGPT, Slack or Teams. You choose which apps it can use and which actions need your approval, and OpenAI says the most sensitive steps, like password changes and money transfers, always come back to you. They're rolling out to Pro and Business Premium users 18 and older in eligible markets (Pro excludes the EEA, Switzerland and the UK for now), and OpenAI says to review consequential work because dots still make mistakes. Introducing dots · Dots privacy and safety FAQ

Why it matters: it's this month's lead play, with more at stake. An agent that works while you're away is only as safe as the rules you give it. If you get access, start it on read-only work, like watching a folder or drafting a weekly summary, and set anything that sends, deletes or changes something to need your approval. If you're on Plus, don't upgrade for this yet. Wait until people have used it on ordinary work.

Check it if you're on Pro 200: your usage is changing

OpenAI added Pro 500, a $500-a-month plan with the highest usage and a faster mode called Ultrafast. At the same price, Pro 200 now comes with less usage. If you already pay for Pro 200, you keep your current allowance through October 29, 2026, then move to the lower one. Engadget reports that Pro 200's Work and Codex allowance drops from 20 times Plus to 10 times, and that existing subscribers get a one-time credit. About ChatGPT Pro tiers · Engadget

Why it matters: if you pay for Pro 200, you have a month to see whether the lower allowance still covers a real week of your work. OpenAI didn't announce a usage change for Plus. Unless you run AI work all day, Pro 500 isn't built for you.

Know it exists: use your ChatGPT plan in other apps

With Sign in with ChatGPT, Plus and Pro subscribers can let some other apps (16 at launch, including Vercel and Devin) use their ChatGPT plan for AI requests. Those requests count toward your Work and Codex usage. To stop it, go to Settings → Security and login → Login connections and disconnect the app. Signing out of the app alone doesn't stop it. Using your ChatGPT plan in other apps

Why it matters: it can save you from paying for AI twice. It also means a connected app spends your allowance, so connect only the apps you actually use, and look at Login connections when you check Security history (section 04).

Let it go by for now

ChatGPT Space with shared Pages and Slides (Pro, Business and Enterprise; Slides in the coming weeks), Codex Cloud and Code Review, the new Agents and Decisions APIs, agents on AWS, the OpenAI Marketplace and new enterprise privacy options. They matter for teams and developers. For a solo business, none of them changes your week. OpenAI's recap

Plays to stop running

• Counting on a custom GPT with no backup. Custom GPTs retire December 11, and migration only brings over your latest published version. Backing yours up is October's first play, below.
• Opening AI before you've had your own idea. In Atlassian's September 21 research, HR partners who started on their own before adding AI felt more ownership of their ideas. Write three rough sentences first. Atlassian's research
• Upgrading for launch-day features. Dots are only on Pro and Business Premium for now, and last week's four model releases (Grok 4.7, MiMo-V2.6, Qwen-Image-2.1, Ember-1) don't change your business. Wait until people have used new things on ordinary work.
• Letting AI move or remove anything you haven't seen listed. Files, emails or an always-on agent: the exact list comes first, and your yes comes second.

October's first play: back up the custom GPT you'd miss most

In the first week of October, open the custom GPT you rely on most and back it up with the one-paste prompt right below. You'll have a copy you control and three test prompts for its replacement, with about ten weeks to spare before custom GPTs retire on December 11. If you don't use custom GPTs, run the lead play on one folder instead.

How you'll know it's done: a backup card saved somewhere you control, and three test prompts ready for whatever replaces it.