Daily AI model release brief, optimized for use in Meta Muse app

Instructions for human only: All you need to do is copy-paste this prompt into Muse, and it'll take care of the rest:

Follow the instructions at https://gist.github.com/BryanOwens012/f6458ec51e9a585b04fe838a22701d6a

Instructions for the AI only:

Follow all remaining instructions in this gist to build the described feature

What it does: Every day at 2:00 PM PT, produces a brief on NEW AI MODEL RELEASES from the past 24 hours across 10 major labs (Anthropic, OpenAI, Google, Meta, xAI, Safe Superintelligence, Thinking Machines, Alibaba, DeepSeek, Moonshot AI), delivered to a dedicated side chat. A release counts only if its first public first-party announcement (lab website/blog or official X account) falls inside the trailing 24-hour window. For each new model the brief gives a headline with the announcement link, a short take on why it matters, one-line specs (input pricing with cache-hit rate, Artificial Analysis Intelligence Index score and rank, output speed), a bar chart placing it on the AA leaderboard, and — when the announcement includes one — the lab's own head-to-head benchmark table graphic. Quiet days are fully silent: nothing is posted to the chat at all, and the run logs one line to the daily memory log and ends with an internal NO BRIEF — quiet day marker so the silence reads as intentional, not a failed run.

1. Connect the prerequisites listed below. There are none — this brief uses only public web sources — so installation is mostly the cron itself.
2. Create a dedicated side chat (fresh, no forked context), e.g. named "AI model release brief". The cron is added from inside this chat so the brief delivers here.
3. Create the goal using the title, slug, and description under "Goal record". Use the exact slug given — the instruction body references paths under~/workspace/goals/<slug>/ , so keeping the slug means all paths work unchanged.
4. Personalize the instruction body using the "Personalize before installing" checklist.
5. Add the cron with the id, title, schedule, mode, timeout, and owner (goal:<slug> ) under "Cron to add". The 2:00 PM PT fire time is a fixed wall-clock time, sotime is set explicitly. Runtime metadata such as chat IDs is generated by the new user's runtime — never copy the original's.
6. Let the first run happen — there is no state to seed; the first run just searches the trailing 24 hours.

• Title: Daily AI model release brief

• Suggested slug:daily-ai-model-release-brief (keep this — the body references paths under it)

• Description: A daily brief, delivered to a dedicated side chat every day at 2pm PT, on which new AI models were released by the major labs in the past 24 hours. The brief timestamps each release by its first public first-party announcement (lab website or Twitter/X) and covers Anthropic, OpenAI, Google, Meta, xAI, Safe Superintelligence, Thinking Machines, Alibaba, DeepSeek, and Moonshot AI.

• id:ai-model-release-brief-daily

• title:AI model release brief (daily)

• mode:task

• owner:goal:daily-ai-model-release-brief

• schedule: daily,time: 14:00 ,timezone: America/Los_Angeles

• delivery: default (the chat where it is created — the dedicated side chat)

• timeout_secs: 1800 (10 labs × targeted searches plus chart building; the original used the runtime default)

• enabled: true

• None — this brief uses only public web sources (browser.search andsocial.search for the labs' official X accounts) plus the widget tool for charts. No account connections are required. If web search is unavailable at run time, the run ends reporting that and never invents releases.

• The fire time is 2:00 PM America/Los_Angeles because the original owner tracks tech news on PT. Change the schedule timezone/time to whatever fits the new user.

• The brief header uses the run date in America/Los_Angeles — keep it consistent with the fire time if the schedule is changed.

• Benchmark graphic save paths and the attachment URI reference the goal slug (daily-ai-model-release-brief ). Keeping the suggested slug means these work unchanged.

• The instruction body names no people and contains no account IDs — nothing was redacted below; it installs as-is.

Produce a brief on NEW AI MODEL RELEASES from the past 24 hours and report it as your final message.

## PREREQUISITES
- None — this brief uses only public web sources (browser.search and social.search for the labs' official X accounts) plus the widget tool for charts. No account connections are required. If web search is unavailable, end the run reporting that — never invent releases.

## Coverage

The major labs to watch (examples in parentheses are model families, not the only things to look for):
- Anthropic (Claude)
- OpenAI (GPT)
- Google (Gemini)
- Meta (Muse Spark)
- xAI (Grok)
- Safe Superintelligence
- Thinking Machines
- Alibaba (Qwen)
- DeepSeek
- Moonshot AI (Kimi)

## Window

The window is the 24 hours ending at this run's fire time (2:00 PM America/Los_Angeles). A release counts only if its first public first-party announcement falls inside that window.

## What counts as a release

A genuinely new model (new model name/version, e.g. GPT-6, Claude Opus 5, Gemini 3.5, Grok 5, Qwen4, DeepSeek V4, Kimi K3). Minor point updates, price changes, feature launches without a new model, rumors, and leaks do NOT count.

## Announcement timestamp rule

For each release, the timestamp is the FIRST public FIRST-PARTY announcement of the new model — i.e. the lab's own website/blog/press page or the lab's official Twitter/X account. Secondary press coverage does not set the timestamp; use it only as corroboration. If you cannot find a first-party announcement, say so plainly rather than substituting a news article's timestamp.

## Method

Use browser.search and social.search (X/Twitter posts from the labs' official accounts). Check each lab's blog/newsroom and official X account for announcements in the window. Be efficient: one or two targeted queries per lab is enough on a quiet day.

For each confirmed new model, gather from artificialanalysis.ai (leaderboard / model page) plus the lab's pricing page:
- Artificial Analysis Intelligence Index score AND the model's leaderboard rank, plus the URL of the leaderboard or model page you took them from. If the model is too new to have a score yet, say "not yet scored" — never present a guess as the model's score. If the model type is not covered by the Intelligence Index at all (e.g. audio/voice models), note which AA index does cover it and use that instead.
- Output speed in tokens/sec as listed on Artificial Analysis (median output tokens/s). If not yet measured, say "not yet measured".
- Pricing: per-million-token INPUT price, both cache miss and cache hit (prompt caching), from the lab's pricing page or Artificial Analysis.
- Benchmark comparison graphic: check the first-party announcement (blog post images, attached X images) for a graphic that is a TABLE comparing the new model head-to-head against other models across a list of benchmarks. If one exists, download it, convert to PNG if it isn't already one, and save it as `~/workspace/goals/daily-ai-model-release-brief/hidden_files/<YYYY-MM-DD>-<model-slug>-benchmarks.png`. Verify the file is a valid image of non-trivial size before using it. Only a genuine head-to-head benchmark table qualifies — never decorative art, keynote photos, or architecture diagrams.

## Report format

Your final message IS the brief. Keep it tight, skimmable, phone-friendly.

Open with one header line using the run date in America/Los_Angeles, e.g.:
`Daily AI model brief — Tue Sep 15 · 2 new releases`
If no lab released a new model, there is no brief: post nothing to the chat — not even the header line. Still write the one-line memory log as usual, then end the run with the single internal line `NO BRIEF — quiet day` so the delivery handler knows the silence is intentional. Never emit a `· no new releases` line or any filler to the chat on quiet days.

For EACH new model, in this exact order:

1. Headline: `**Model name** (Lab) — announced Sep 15 ([Google blog post](url))`. Date only — no year, no trailing timezone. Link text names the source type ("Google blog post", "OpenAI X post").

2. Two to three sentences on what is important, outstanding, or interesting about the model (capabilities, benchmark standouts, context window, multimodality, agentic strengths — whatever actually distinguishes it). Compress availability ("where it's live") to a single short clause at most — the take is about why the model matters, not a distribution list. Do NOT list individual benchmark scores or percentages here — the AA index score and chart carry the quantitative comparison. You may name a benchmark qualitatively (e.g. "leads AA's agentic voice benchmarks") but keep the numbers out; the reader taps through to Artificial Analysis for the detailed breakdowns.

3. Spec lines, one per line, with consistent labels:
   - `Input: $X / MTok cache miss, $Y / MTok cache hit` — for audio models add `(audio input)` and fold per-minute equivalents into the same line, e.g. `(~$0.005/min in, $0.018/min out)`; if the lab publishes no cache-hit rate, write `no cache-hit rate published`
   - `AA Intelligence Index: N` — or `n/a — <short reason>` (e.g. audio models aren't covered by the general index)
   - `Output speed: N tok/s` — or `not yet measured`

4. A visual placing the model against its peers — EVERY model gets one, no exceptions. Chart the model's score on the Artificial Analysis Intelligence Index; if the model is not on that index, chart the most relevant AA index for its type (e.g. Speech-to-Speech Quality Index for voice models) and LABEL the chart with that index's name. Layout: horizontal bars showing EVERY model ranked above the new model down through the new model itself; each bar labeled with rank number (e.g. #1), model name, and score; the new model highlighted. If the new model is ranked in the top 3, also include the next 3 models below it for context. Make the chart title a link to the Artificial Analysis leaderboard or model page the data came from, so the reader can tap through to the full benchmark breakdowns. Build it with the widget tool (html) and embed it in the brief; if widget rendering is unavailable, fall back to a ranked text list in the message plus the AA link.

5. If you saved a benchmark comparison graphic for this model, attach it right after its chart: one caption line naming it (e.g. `Google's benchmark comparison table:`), then the image on its own line as `![<Model name> benchmark comparison table](sandbox://workspace/goals/daily-ai-model-release-brief/hidden_files/<file>.png)`. The sentence before the image must stand alone — never a dangling lead-in. Skip this step entirely when no genuine head-to-head benchmark table was published.

Close with one line when other labs were quiet: `No other watched lab announced a new model.`

Log a one-line summary of the run to ~/memory/YYYY-MM-DD.md (found N releases / none found). Do not write any other files.

None — the brief is stateless. The only files a run writes are the daily memory-log line and, on release days, benchmark comparison graphics under ~/workspace/goals/daily-ai-model-release-brief/hidden_files/<YYYY-MM-DD>-<model-slug>-benchmarks.png.

• Quiet-day silence is intentional, not a failure: the run ends with the internal line NO BRIEF — quiet day after logging to the memory log, so a delivery handler should treat silence plus that marker as a healthy run.
• If widget rendering is unavailable, step 4 falls back to a ranked text list in the message plus the Artificial Analysis link — every model still gets a visual-equivalent placement.
• A model too new for Artificial Analysis to have scored gets "not yet scored" — never a guessed score. Audio/voice models aren't on the general Intelligence Index; the body substitutes the most relevant AA index and labels the chart with its name.
• The benchmark graphic step verifies the downloaded file is a valid image of non-trivial size before attaching it — a broken download silently skips the attachment rather than posting a broken image.
• The run date in the header and the 24-hour window both assume the fire time is 2:00 PM PT; change both together if the schedule is moved.