SI Agent News

• Feature request: a user asks to @mention an existing Dot from any ChatGPT conversation. CONFIRMED that the post exists; it is a user request, not an OpenAI plan. OpenAI community forum
• Feature request: GPTGIFS, custom AI GIFs and community reactions built into ChatGPT. CONFIRMED as a post; STILL OPEN as anything more. OpenAI community forum
• Forum post: a user trying to reach the ChatGPT product design team. No product news. OpenAI community forum

Takeaway: by the look of the first title, people want to reuse something they've already set up instead of starting over.

ChatGPT reportedly gets a team feature for assigning and tracking tasks

36Kr, a Chinese tech news outlet, reports that ChatGPT (OpenAI's AI assistant app) is getting a team feature. You can assign tasks separately to different members and track progress at any time. The headline also mentions a 'GPT-6 Astra Pro' upgrade, but we have no detail on what it changes.

Status: the team task feature is REPORTED: one outlet, no OpenAI release note in hand. The 'Astra Pro' name comes from a single aggregator source, so it is STILL OPEN. Price: not stated. Worth watching if you coordinate a small group with recurring tasks. If you work alone, the lite version in today's technique works now. Until OpenAI confirms anything, don't rebuild your workflow around this.

What this means for you:

• Small business owner: weekly handoffs, like who sends the quote, can live in one place instead of scattered messages.
• Job seeker or student: a group project or job search is a task list. Practice writing each task with an owner and a done-when line.
• Creator: a video is a chain of handoffs: script, edit, thumbnail. Tracking each step is where this feature would earn its keep.

AI reverse engineering reportedly replicates Adobe and Microsoft software

The Chosun Ilbo, a Korean newspaper, carried this story on MSN under the headline 'AI reverse engineering replicates Adobe, Microsoft software.' Reverse engineering means studying how a finished program behaves and rebuilding something that works the same way. We have only the headline, so how complete the copies are, which products were copied and who did it are all STILL OPEN. Status: REPORTED, thinly.

Why it might matter to you is our analysis, not the report's. If AI makes look-alike tools easier to build, cheaper alternatives to familiar paid software may appear. So might unvetted copies. Practical moves: keep your important work in widely supported file formats so you aren't locked to one vendor. Download software only from the maker's own site. Before using any look-alike for client or paid work, check its license terms. This matters most to creators and small businesses who pay monthly for design or office software.

Chinese AI firms challenge U.S. rivals on price

Dong-A Ilbo, a Korean newspaper, reports the price challenge. REPORTED, thinly: we have no figures and no company names, so any detail is STILL OPEN. Competition on price usually pushes prices down; that is our analysis, not the report's.

Practical move: keep your prompts in plain files you own so switching providers is easy, and check each company's pricing page about once a quarter. Worth it most for people paying per use through an API (pay-by-the-amount access for developers). If you pay a flat monthly fee, the effect is less direct. Still, a cheaper rival can change what the bigger companies include in their plans. When you compare options, test the same few real tasks on each one, not just the price list, because a cheaper tool that needs more rework isn't cheaper.

Community request: @mention an existing Dot from any ChatGPT conversation

A post on community.openai.com, OpenAI's user forum, is titled 'Feature request: Let me @mention an existing Dot from any ChatGPT conversation.' We have only the title. 'Dot' is the poster's term, and we haven't confirmed what it refers to in ChatGPT. Status: a user wish, not an announced feature, and there is no sign OpenAI has responded.

The underlying idea is useful anyway: pull something you already set up into a new chat by naming it, instead of re-explaining it each time. Whatever the feature turns out to be, you can approximate that habit today. Keep a short note with your standing instructions, such as your business voice, your résumé facts or your project background, and paste it at the start of a new conversation. This helps anyone who keeps restarting chats and retyping context, especially small business owners and students juggling several projects.

Community request: GPTGIFS, custom AI GIFs and community reactions

Another forum post, 'Feature Request: GPTGIFS — Custom AI GIFs and Community Reactions Built Into ChatGPT,' asks for custom AI-made GIFs and shared reactions inside ChatGPT. Again we have only the title, so details are STILL OPEN. It is a user suggestion, not a product plan, and OpenAI hasn't said it is considering it.

It is the least practical item today, but it shows what some users want: more expressive, shareable output from a chat tool. If you are a creator, don't wait for it. Short animated reactions are a common way to get attention on social posts, and existing tools can already make them. If you try any, check each tool's rules on commercial use before you post for a brand or client. Everyone else can safely skip this one.

Community post: trying to contact the ChatGPT product design team

A third forum thread is titled 'Trying to contact someone on the ChatGPT product design team.' We don't know who posted it, what they need, or whether anyone replied. Status: REPORTED only as a title, and the outcome is STILL OPEN.

The useful takeaway is about how feedback reaches a company. The forum is full of feature requests, like the two above, and the poster is evidently looking for a more direct route. Our analysis: if you want a product change, a clear, specific post is more likely to be useful than a vague complaint. State what you were trying to do, what happened, and what you wish had happened. Use the official help or feedback channels for account problems, because a public forum thread isn't a reliable way to reach staff. Be wary of anyone who privately offers to put you in touch with OpenAI staff.

AI reverse engineering replicates Adobe and Microsoft software, report says

The Chosun Ilbo, a Korean newspaper, via MSN, carries the headline. That is nearly all we have. REPORTED at headline level; which products, how close the copies are and who is doing it are STILL OPEN. 'Replicates' could mean a convincing demo or a working product, and those are very different.

What it means (our analysis, not the report's): reverse engineering means studying a finished product's behavior, then rebuilding something that does the same job. It is old. The new part is AI making the studying and rebuilding faster. If copying behavior gets cheaper, a feature list is a weak moat. The stronger protection is likely data and file formats people already depend on, distribution, trust with large customers, and integrations.

Who benefits: buyers with leverage, and small teams wanting a cheaper tool for one narrow job. Who is squeezed: incumbents' pricing power on simple tools.

Legal note: in many places copyright protects code and expression, not the idea of what a program does, but details vary by country and case. This is not legal advice.

Do this: if you buy software, ask vendors what they do that a copy couldn't. If you build, don't assume your features are a moat. If you use AI to rebuild something, check the license of whatever you studied first.

What would settle it: a named, testable example, a response from the companies, or a legal claim. None is in what we have.

A lesson, not news. This explains how OpenAI's developer tools generally work; it does not describe how ChatGPT's team feature is built, which we don't know.

For builders: the mechanism is tool calling (also called function calling).

1. You give the model a list of tools. Each has a name, a plain-language description, and a schema: a template saying what inputs it expects.
2. When the model decides a tool is needed, it runs nothing. It outputs a structured request: this tool, these arguments.
3. Your code runs the tool and sends the result back. The model continues, maybe asking for another tool, until it writes a final answer. It is a loop.

Task tracker example: three tools: list tasks, assign a task, update a status. The key design choice is that tasks live in your database, not in the chat. The chat is a window, which is why it can answer 'what's blocked?' accurately days later.

Why it works: models are trained to produce output that fits a schema, and your descriptions act as instructions, so the tool descriptions are prompts. Vague descriptions give vague calls. OpenAI also offers a strict structured-output option that forces arguments to match the schema; check the current documentation.

Novel or incremental? Incremental. The loop is not new; the packaging is.

Limits: arguments can still be wrong, so validate before anything that writes or deletes. Give each tool the least permission it needs. Every trip around the loop resends the conversation, so long loops cost more. Cap the number of steps so loops can't spin. You can also test each tool alone, before the model is involved.

Non-builder takeaway: an assistant tracking tasks is only as good as the list it reads from.

The task ledger. Illustrative workflow, not tested by us. It works in ChatGPT or any assistant. For anyone juggling more than about five open tasks; skip it if you have fewer.

1. Write the goal in one sentence.
2. Break it into tasks. Give each an owner, a done-when line, a due date and any input.
3. Paste the ledger into a new chat with this rule: 'This ledger is the single source of truth. After every update, return the full ledger.'
4. Each day, paste your messy notes and ask for two things only: what changed, and what's blocked.
5. On Friday, ask for a status summary written for whoever you report to.

Task 3: Draft client quote
Owner: Sam
Done when: quote sent and logged
Due: Thursday
Input: price list
Status: in progress

Limits: a chat's memory is not a database, so re-paste the ledger at the start of every new chat. The assistant can quietly drop or reword a row.

How to check it worked: count tasks before and after each update. The number should change only when you meant it to. Scan statuses against your own notes.

About 15 minutes a week. Suppose your weekly status update takes about 20 minutes by hand, and with the ledger it takes about 5, mostly review. That saves roughly 15 minutes a week, about 13 hours a year, or around a day and a half of work. This is our estimate, not a measured result.

It isn't tied to one vendor: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Copilot (Microsoft) can all work from a pasted ledger. Today's stories state no prices, so check the current price and limits on each pricing page.

Second saver: don't pay for a new team tool on a report. The team feature is reported, price not stated. Run the ledger for a week and you'll know what you would actually want.

What you give up: the 5 minutes is review time. Skip it and an error goes straight into your report.

Prompt clinic: checking AI pricing.

Weak: 'Which AI is cheapest?' It doesn't know your workload and may answer from memory with outdated prices.

Better:

I will paste pricing text copied from official pricing pages. Use only the numbers in that text. My workload: [describe it, for example 200 support replies a day, about 300 words each]. For each option, show a table with the price unit, what it would cost for my workload per month, and your formula. Flag anything missing instead of guessing. End with the two options that look cheapest and what I would give up with each.

Why it works: it supplies the inputs, limits the source, shows the formula so you can check the math, and allows 'missing'.

Expected result: a small table you can check line by line. Customize: change the workload; add a quality bar such as 'must handle documents longer than ten pages'. Quality check: recompute one row by hand. If it's off, don't trust the rest.

Not from today's news. In 'Lost in the Middle' (Nelson Liu and colleagues), models did best when the relevant information sat at the start or end of a long input, and worse when it was buried in the middle. A context window is the amount of text a model reads at once. Caveat: the models tested were from that period, so newer ones may do better.

Decision this week: in long pastes, put your instructions and the most important item at the top, restate the question at the end, and test it on your own task.

Agentic AI, meaning assistants that carry out multi-step work, is the busiest topic today. Today's OpenAI items fit: task tracking, price pressure, copied software.

Watch: whether OpenAI itself confirms the team feature and the 'Astra Pro' name (no date announced), and whether real price figures appear behind the China story. Either would confirm or end those reports.

I asked my AI assistant to track my tasks. It assigned itself a vacation and marked it done.

That's it for today. Tomorrow we're watching whether OpenAI confirms the ChatGPT team feature, and whether real price figures appear behind the China story.

If this made you smarter, share it with one person who needs it. SI Agent News is independent and self-funded, with no investors and no ads., and every share and every bit of support keeps it going. Thank you.

Made by AI.