ChatGPT can read and send your messages now.
Keep the final check.
This was a week when AI tools reached further into real work.
The useful question is not what they can touch.
It is whether they know where to stop.
Start with the useful partGet the seven decisions that matter this week.
See the whole weekRead everything I kept.
What I read. What you can try. What I left alone.
My goal: save you time, reduce noise, and help you make steady progress.
Overwhelm beat
More access does not automatically mean less work
The strongest member signal this week was not wanting another tool.
It was wanting fewer open loops, less repeated explaining, and one finished thing.
That matters because the week's AI news moved in the opposite direction on the surface. More tools can read your files, work in your browser, use your messages, and take actions for you.
That can remove real work. It can also create another place where you need to check permissions, verify context, correct a draft, and make sure the tool did not go further than you intended.
A handoff reduces cognitive burden when the boundary is clear before the task begins.
One source. One output. One approval. One place where the AI has to stop and ask.
That is enough structure to finish something without building another system around it.
Read this first
This issue in 30 seconds
ChatGPT can work inside Apple Messages.
Start with one draft on an Apple Silicon Mac. Keep approval on for every send.
Some older ChatGPT Zaps stop August 26.
Look only if you already use the named legacy Assistants actions.
Claude Academy is open.
Choose one lesson tied to work you already need to do. Do not collect another curriculum.
Replit Free Mode can stretch your plan.
Put ordinary planning and small fixes in Free Mode. Save paid credits for work that needs more power.
College students can claim a year of Google AI.
Set a cancellation reminder when you claim it because the plan renews automatically.
ChatGPT for Teens is rolling out.
Review the learning experience and controls together instead of treating defaults as supervision.
Anthropic's agent APIs are generally available.
This matters to builders. It is not a new consumer Claude feature.
If you do one thing: take one repeated task and name one source, one output, one approval, and one stop condition.
ChatGPT can work inside Apple Messages now
The best first test is still a draft you approve.
OpenAI added an Apple Messages plugin to the ChatGPT desktop app for Apple Silicon Macs on August 20. It is available on all ChatGPT plans, but the Messages tools work only in ChatGPT Work and Codex.
The plugin can read and search iMessage, SMS, and RCS conversations on your Mac. It can also prepare or send messages through the Messages app.
It does not work in a regular ChatGPT chat, on the web or mobile, in Codex CLI, or in the IDE extension.
The default is the part I care about: ChatGPT asks you to approve the message and its recipients before it sends.
You can always allow sending to a specific chat. I would not do that during a first test. Persistent approval removes the last chance to catch the wrong recipient, an invented detail, or a message that sounds nothing like you.
Where this could help
- Find the latest question in a long logistics thread.
- Draft a reply when you know the answer but keep postponing the wording.
- Pull dates or decisions from a conversation before updating a real source document.
- Prepare a short follow-up without hunting through the thread yourself.
Where I would not start
- Client, employee, health, financial, legal, or password-related conversations.
- A message that makes a commitment or changes a relationship.
- A group thread where the correct recipient is not obvious.
- Any chat where persistent sending approval is already on.
My decision: Use it for one ordinary draft. Read it. Send it yourself or approve that one message. Make the tool earn anything broader.
Some older ChatGPT Zaps stop working August 26
This is urgent for a small group and safely ignored by everyone else.
Zapier says Zaps using several deprecated ChatGPT Assistants API actions will stop working August 26.
The affected actions include Conversation With Assistant (Legacy), Create Assistant, Upload File when used with assistants, Find Assistant, and Find or Create Assistant.
Zapier will automatically migrate the legacy Conversation action, but it leaves the migrated Zap turned off so you can review it. The other affected actions need to be rebuilt or redesigned.
If those names mean nothing to you, ignore this story.
If you use older ChatGPT Zaps, search your Zaps for ChatGPT, open each one, and check the action name. Test the replacement and confirm downstream fields before turning it back on.
Claude Academy gives you a better place to learn than the feature feed
Pick one lesson that answers a question already slowing you down.
Anthropic launched Claude Academy on August 20. It includes product tutorials, role-based use cases, and model-agnostic AI fundamentals.
The part I like is the focus on decisions that survive a product update: what should be delegated, what should stay with you, what the AI tends to get wrong, and how much verification the stakes require.
Claude Academy currently includes a four-hour AI Fluency course, shorter tutorials, and learning paths for Claude, Cowork, Claude Code, Claude Tag, and the Claude Platform.
Do not try to complete it all. The seven-minute tutorial on the four properties of AI is a better start than opening five courses and finishing none.
Replit Free Mode can stretch the plan you already pay for
Use the lighter mode for ordinary work. Save stronger modes for the build that needs them.
Replit launched Free Mode on August 18 for Core and Pro subscribers. It runs on OpenAI's GPT-5.6 Luna and does not use credits for everyday chat and Agent tasks while you are within the mode's usage limits.
Those limits reset every five hours. Replit says Pro members receive more Free Mode usage than Core members.
Free Mode is not unlimited, and it is not the same as a free Replit account. For more complex work, Replit may suggest Power or Max Mode, which uses paid credits.
If you already pay for Replit, use Free Mode for brainstorming, planning, copy, and small fixes. Move up only when the task needs it.
If you do not already use Replit, this is not a reason to add another tool. The useful change is for people who were already watching credits disappear during routine work.
My decision: Use it if it appears in the plan you already have. Keep ordinary work in Free Mode and save paid credits for builds that need a stronger mode.
College students can claim one year of Google's paid AI plans
The free year comes with an automatic renewal, so set the reminder now.
Eligible college students in the U.S. can claim one year of Google AI Pro at no cost. Eligible students in more than 140 other markets can claim one year of Google AI Plus where the offer is available.
In the U.S., AI Pro includes higher Gemini limits, Gemini Spark, Gemini inside apps such as Gmail and Docs, and 5 TB of storage. Google values the plan at $19.99 per month.
A valid payment method is required. The U.S. plan automatically renews at $19.99 per month after the year unless you cancel. The AI Plus offer renews at $4.99 per month or the local equivalent.
The claim window runs through December 31, 2026. If you qualify and would use it, claim it and immediately add a calendar reminder for eleven months from now.
ChatGPT created a separate teen experience
A safer default is useful. It is not the same as supervision.
ChatGPT for Teens began rolling out globally August 18 to eligible teen accounts on Free and paid personal plans.
OpenAI says accounts are placed into the teen experience when account information, verified age, or age prediction indicates the user is under 18. The experience includes Study Mode, step-by-step learning support, homework reminders, sensitive-image warnings, stronger content safeguards, and optional linked-parent controls.
Parents with linked accounts can set quiet hours and manage selected settings. OpenAI says safety notifications are limited to certain higher-risk situations and share only the information needed to support the teen.
This does not guarantee that age prediction is always right, and it does not turn an AI tool into a teacher, parent, or mental-health professional.
If a teen in your life uses ChatGPT, look at the experience together. Ask it to explain one problem step by step, then talk about what it got right, what still needs checking, and what should stay between the teen and a real person.
Anthropic made its agent-building layer more production-ready
This matters to builders. It is not a new feature inside a normal Claude subscription.
Anthropic made computer use, a new browser use tool, the Skills API, and the Files API generally available on the Claude Platform on August 20.
Together, those pieces let developers build an agent that can operate software, follow versioned instructions, reuse uploaded files, and return finished documents.
This does not mean every Claude subscriber suddenly has a new autonomous agent inside the normal Claude app. It means teams building on Anthropic's API have a more complete set of building blocks.
For most readers, watch what this changes in tools you already pay for. Do not connect more accounts because the underlying APIs became generally available.
For builders, the design question stays the same: what is the smallest job, what source can it use, what action requires review, and what evidence proves the task is done?
My take
The feature is not the finished workflow
The tools are getting better at reaching the work. That does not mean you need to give them everything.
Messages is useful when it saves the search and gets a reply out of your head. It is less useful when you have to monitor a system with broad permission to speak as you.
The same thing is true for files, browsers, inboxes, and agent systems.
If you cannot say what source the AI should use, what output it should make, and what still requires your approval, the workflow is not ready to run without you.
Before you add an AI-assisted step, ask what existing step, report, or recurring task it will replace. If nothing leaves, the workflow may be adding work instead of removing it.
You do not need a complete automation plan. You need one handoff small enough to review and useful enough to repeat.
Put it to work
The ten-minute safe handoff
Use this on Messages if you have the new plugin. Use the same structure in email, a document, or an ordinary AI chat if you do not.
Use this when
You have one repeated task and one source that contains enough information for a reviewable draft.
Skip it when
The task includes confidential data, an unclear recipient, a consequential decision, or an action you cannot reverse.
Choose one conversation, document, spreadsheet, or folder. Do not hand over an entire inbox or drive.
Ask for one draft, checklist, comparison, or update that you can review in a few minutes.
Say what the AI may prepare and what it may not send, publish, delete, change, or promise.
Tell it when uncertainty should end the task and bring the decision back to you.
If no step becomes simpler, shorter, or unnecessary, the AI may be adding another checkpoint instead of reducing work.
Use only [SOURCE] to prepare [OUTPUT]. This handoff replaces [OLD STEP]. Do not send, publish, delete, change the source, or make commitments for me. Stop and ask me if the source is incomplete, the recipient is unclear, or the task requires a decision I have not made. Show me the source details you used and the finished draft before I approve anything.
Done means: One useful draft exists, you know where every fact came from, nothing happened without your review, and one old step is smaller or gone.
Free resource
The Safe AI Handoff Card
Turn one repeated task into a bounded handoff with one source, one output, one approval, one stop condition, one proof of completion, and one old step to remove.
One week later
What Computer History has actually helped me do
Last week's brief was based on less than 24 hours of using Computer History. After a full week, the useful part is clearer.
I have not used it much to recover one specific lost click. I have used it to ask better questions about the patterns across my work:
- Where am I still doing work AI should handle?
- Where do I keep getting close to finished and then stall before publishing?
- Which things I repeat should become reusable skills instead of another task I remember how to do?
That produced more than an interesting personality report. Computer History surfaced repeated work around member bug reports, Microsoft Forms redesigns, and state-manual comparisons. I turned those into three real skills I can call again instead of rebuilding the process each time.
It also caught a bigger problem. I am very good at using AI to research, analyze, and make plans, then I still manually carry too much of the result across tools and through the final checks. That finding pushed me toward a Finish-Line Agent and a simpler work-control system built around what is active, waiting, ready to verify, and actually done.
The feature did not solve my attention problem. It made the pattern hard to ignore.
My one-week verdict: Keep using it for weekly reflection, finding repeatable work, and noticing where I am creating more analysis instead of finishing. Do not treat its summaries as perfect, do not give it decision-making authority, and tighten what it can observe before sensitive work ever appears in the trail.
OpenAI also expanded Computer History to Pro users in the EEA, United Kingdom, and Switzerland on August 20. It remains off by default and requires Memories. I corrected the availability line in last week's guide.
No homework required
Safe to leave alone
GPT-5.6 Sol API price cut
Useful for developers paying by the token. It did not lower a ChatGPT subscription price.
Claude Security with Mythos 5
An Enterprise code-security beta and partner program. It still requires human review of proposed fixes.
ChatGPT Ads expanding in Europe August 24
Free and Go users in 31 European markets may see ads. Plus, Pro, and Enterprise remain ad-free. Self-service Ads Manager is not open yet, so most members have no advertising task today.
Another round of model comparisons
Several models released or changed just outside this issue window. None gives you a permanent winner or a reason to rebuild a working stack this week.
Source note: Product, plan, availability, and deadline claims were checked against current official pages on August 23. The member framing uses anonymous aggregate signals only.
What I'm watching
Whether the new action surfaces actually reduce repeat work after the novelty wears off, and which products keep review easy enough that people can still tell what happened.
One question
What is one thing you would let AI prepare this week, but still want to approve yourself?