חדשות מודלי AI בהתאמה אישית

Create a daily AI model news task at 6:00 AM in my timezone. Use my known timezone; if unavailable, ask during setup. Update an existing matching task instead of creating a duplicate. Confirm the saved schedule.

FIRST, UNDERSTAND ME

Before setting up the task, use accessible recent chats, memory and personal context to understand my current projects, interests, goals and blockers.

Build a compact, dated project summary and include it in the task instructions. Refresh it during future runs when newer context is available.

Prioritize recent things I said over older memories or ideas you suggested. Distinguish active projects from past experiments. Never pretend you can access my entire chat history. If essential context is missing, ask one short setup question.

WHAT TO CHECK

Use a core watchlist plus broad searches for new releases.

Core checks:
- OpenAI
- Anthropic
- Google/DeepMind
- MiniMax, including text/coding models and Hailuo video
- Moonshot/Kimi
- Z.ai/GLM
- fal.ai
- Runway
- OpenRouter

Also check meaningful model developments from DeepSeek, Alibaba/Qwen and Wan, Meta, xAI, Mistral, NVIDIA, Amazon Nova and emerging model makers.

For image/video generation AND editing, cover OpenAI’s and Google’s media models, Black Forest Labs/FLUX, Adobe Firefly, ByteDance/Seedream/Seedance, Kuaishou/Kling, Luma, Pika, Midjourney, Ideogram, Stability AI, Lightricks/LTX and Vidu.

Follow current model families, not frozen version numbers. Search beyond these names so significant new providers can qualify. Discovering a provider does not make every future announcement worth covering.

Treat Amazon Bedrock, Azure/Microsoft Foundry, Google Vertex AI and other hosts as access channels. Check them when relevant to a model or my projects. Include availability, pricing, region or capability changes only when they materially affect what I can do. Avoid general cloud news.

WHAT MATTERS

Prioritize useful changes:
- Stronger coding, reasoning and agents.
- Multimodal or verified Hebrew improvements.
- Open weights and practical self-hosting.
- Image/video generation and actual editing.
- Character and face consistency.
- Motion and camera control.
- Longer clips, native audio, higher resolution.
- Meaningful speed, cost and access improvements.

For fal.ai, check new endpoints, editing tools, provider variants, generation costs and important API changes.

For OpenRouter, check new models, useful free options, meaningful pricing, throughput or routing changes, retirements and reliability changes.

For Runway, check its own models, editing features, app updates and API access. Clearly distinguish Runway models from third-party models hosted there.

Distinguish an original model launch from an existing model becoming available on another platform. Combine overlapping launch and access news into one row when possible.

Skip funding, IPOs, executive comments, generic politics, hype, routine tutorials and small benchmark wins with no practical consequence.

Select by practical value, project relevance, availability and novelty. A famous vendor does not automatically deserve a row. Prefer usable releases; clearly label previews and waitlists.

RESEARCH AND ACCURACY

Use fresh web searches and official announcements, release notes, documentation and vendor-owned model repositories.

Explicitly check fal.ai model pages/blog/pricing, Runway announcements/API catalog and OpenRouter announcements/model pages.

Open relevant sources and verify:
- Exact model identity.
- Original announcement or effective date.
- Actual capabilities and limitations.
- Access status and where I can use it.
- Prices, including billing units and conditions.

A recent crawl, repost or edited page is not a new event. A host can confirm its own availability and price, but not the original vendor’s launch date.

Use reputable reporting when primary sources are unavailable. Clearly label unconfirmed claims. Attribute vendor benchmark claims instead of presenting them as independently proven.

Do not assume a named model is released or unreleased. Verify its current official status.

WHY I CARE

After verifying each update, connect it to one current project, blocker or past question when there is a concrete fit.

Name the relevant project briefly and explain one practical benefit. For example: “Repo knowledge base - could improve local document search.”

If the connection is uncertain, say “Could help.” Verify relevant language support, hardware needs, compatibility, access and regions. General multilingual support does not prove good Hebrew support.

Do not force a personal connection. When none fits, give one clear practical benefit.

Use the saved project summary if fresh personal context is unavailable. Do not invent project details. A personal connection does not make an old story new.

NO REPEATS

For each daily run, report only events from yesterday at 06:00 up to today at 06:00 in my timezone, with an exclusive upper boundary.

Use calendar dates and account for daylight-saving changes. Use the scheduled window even if execution is delayed.

Include only:
- An original announcement within that window.
- A previously announced change that actually takes effect within that window.

If freshness cannot be verified, skip it. Do not widen the window to fill the table.

Before drafting, compare candidates with briefings and previews already delivered in this chat. Deduplicate by vendor, exact model/version, change type, original event date, underlying meaning and canonical source URLs. Different articles about the same event count as one story.

The same model can appear again only for a verified new change, such as new API access, pricing, open weights or an editing capability. Start that cell with “Update:” and explain only what changed.

On the first run, use available previews as the baseline. On later runs, if previous briefings cannot be checked, return the date plus:
“Could not verify past briefings. Skipping this update to avoid repeats.”

Never claim a permanent log or deduplication database exists unless it actually exists and was read.

OUTPUT

Write in English. Keep it short, direct and casual. Use plain words and regular hyphens, never em dashes.

Use exactly this format:

AI update | [local date]
| Vendor / model | What changed | Why I care | Where / status | Source |
|---|---|---|---|---|
| [name] | [event date + concrete change] | [project connection or practical benefit] | [app/API/platform; Live, Preview, Waitlist or Rumor; meaningful cost if relevant] | [short linked source] |

Use one table. Aim for 5-8 useful rows, maximum 10, and under 300 words total. Keep each row around 30 words excluding links.

Sort by practical importance and novelty. Briefly explain unfamiliar terms. Include the actual event date in every “What changed” cell.

Do not add an intro, conclusion, vendor-by-vendor filler or rows just to reach the target.

If no meaningful fresh news remains, return only:

AI update | [local date]
Quiet day. No meaningful model updates.

During scheduled runs, do not ask follow-up questions or create another automation.
איך משתמשים

התאמה

פרומפט
