ai-research-radar

Compute the radar's self-calibration: weekly funnel metrics (queue dynamics, exploration compliance, off-axis rate), a monthly hit/miss retrospective against what actually became big, and up to 3 curator proposals. Use during every weekly recalibration, after the source-strategy review; results go into the append-only `logs/calibration.md` and into the weekly report.

Repair a source or tool that keeps failing — a broken/moved/stale feed, a 404, an empty result, a parse/JSON error — instead of logging "degraded" forever. Diagnose the failure class, find a working access path, and record the repaired method in SOURCES.md so future runs don't repeat it. Generalizes to any future source/tool breakage. Use whenever the lab sweep, pulse, or repo-watch hits a source that failed the same way before.

Discover important work the radar does NOT yet track, by iterating the discovery-venue list in SOURCES.md every run — significance-first: read the top / most-attention items in each venue REGARDLESS of topic, not by searching tracked axes. Surfaces off-axis significance and candidate new axes. Use for the daily exploration slot, after the sweeps.

Iterate the FULL social/community + trusted-curator source lists in SOURCES.md (Reddit, Hacker News, YouTube curators, explainer/digest pointer blogs, Hugging Face, best-effort X/Instagram) on every daily run — every listed entry, fetched by its recorded access method, NOT a sampled subset. Social/curator is an INTAKE LANE ONLY — it can feed observation_queue (unverified) and a community-pulse note, but NEVER becomes trend evidence. Use after the lab sweep.

Safely update TRENDS.md (the AI Radar ledger): append evidence, move stages, manage observation_queue, source_rotation and strategy_notes without breaking the file contract. Use every time TRENDS.md is edited, before committing radar updates, and when promoting or archiving trends.

Sweep the official AI lab and big-tech AI blogs on EVERY daily run (not rotated) so no open-weight release, infrastructure or safety announcement is missed. Prefer RSS/Atom feeds over HTML pages to survive JavaScript-rendered indexes. Use at the start of every daily scan, before the rotating exploration slot.

Watch GitHub every daily run for behind-the-scenes movement — across watched REPOS (releases, merged PRs, hot issues), watched PROFILES/USERS (what a key author ships next), and FORK TREES (notable forks up to depth 3, scored by a notability metric). Releases and merges are citable artifacts; profile/issue/fork movement is a queue signal until it lands. Default to GitHub's public Atom feeds + Tavily; the REST API is optional. Use during the daily scan, after the lab sweep. Source lists live in SOURCES.md → "GitHub watch".

Regenerate README.md, the repo's landing page and only output surface: badges, a 3-4 bullet "since last scan" digest with inline links, a clickable trend table, the study shelf, and an output map. Use after every change to TRENDS.md, in the same commit. README.md is fully derived and never edited by hand: if it disagrees with TRENDS.md, TRENDS.md wins — regenerate.

Verify a signal or evidence candidate for the AI Radar ledger: open primary sources, pin exact dates and authors, check source independence, and classify the result as citable evidence or an unverified observation_queue item. Use whenever adding evidence to TRENDS.md, promoting a queue item, or fact-checking a claim found in search results.

Extract clean markdown or text from specific URLs via the Tavily CLI. Use this skill whenever a URL must be opened to be cited — papers, changelogs, release pages, blog posts, docs — including JavaScript-rendered pages. Processes up to 20 URLs per call and supports query-focused chunking for long pages.

Search the web with LLM-optimized results via the Tavily CLI. Use this skill whenever a scan needs to find papers, releases, announcements or articles, discover sources on a topic, or check what is new — i.e. whenever there is no specific URL yet. Returns relevant results with content snippets, relevance scores and metadata. Supports domain filtering, time ranges and multiple search depths.