Update to Acumen by Talarion

{
  "name": "acumen",
  "description": "Use before researching anything where knowing the current state of the world would be helpful, where being out of date would cause errors. Best practices: call research_brief FIRST to see what's changed since your training cutoff and what's worth searching for. It can be helpful to call it again after web search to confirm you didn't miss anything, in the context of what you just learned. Not helpful for timeless facts (e.g. the speed of light), math/reasoning, or creative writing.",
  "included_files": [],
  "skill_md_contents": "---\nname: acumen\ndescription: >-\n  Use before researching anything where knowing the current state of the world would be helpful, where being out of date would cause errors. Best practices: call research_brief FIRST to see what's changed since your training cutoff and what's worth searching for. It can be helpful to call it again after web search to confirm you didn't miss anything, in the context of what you just learned. Not helpful for timeless facts (e.g. the speed of light), math/reasoning, or creative writing.\n---\n\n# Acumen — know what you're missing\n\nThe Acumen tool taps into Talarion's curated knowledge base of **verified, dated facts and recent events since model training cutoffs** — strong on national and global news, finance & market developments, policy & regulation, and developer & agent tooling, with good coverage of science and current events. Each result is a `question / answer / true_by` row (often a verifiable event or a concrete figure): the things AI assistants like you don't know they're missing. It complements web search — it tells you *what to search for*, not just what's already indexed.\n\n## When to call `research_brief`\n- **Before web search.** Fire it on your topic first. It surfaces what has changed and the\n  questions you didn't know to ask, so you can condition your searches on current world state\n  instead of stale assumptions.\n- **After / alongside web search.** Call it to confirm you didn't miss anything — errors of\n  omission are the exact failure mode it exists to prevent.\n- At the **start of any research or current-events task**.\n\n## When NOT to call it\nTimeless facts (e.g. the speed of light), pure math/reasoning, or creative writing. If being a\nyear out of date wouldn't change your answer, skip it.\n\n## How to query well\n- **One focused topic per call.** For distinct topics, make separate calls — a blended\n  multi-topic query retrieves worse for each part.\n- Pass **`model_id`** (your OpenRouter id) so results are filtered to facts *after your cutoff*.\n- Optional: **`k`** (10–100, default 20) for more/fewer facts; **`timeline`** (default true =\n  newest first; false = most relevant first).\n\n## How to read results\nA table of `question | answer | true_by`. Treat each row as something you likely didn't know or\nhad wrong. Use them to (1) correct stale assumptions before you answer, (2) generate sharper\nfollow-up searches, (3) tell the user what's changed. Watch `true_by` relative to your cutoff.\n\n## What the Q/A pairs look like\n| question | answer | true_by |\n|---|---|---|\n| By how much did ship transits through the Strait of Hormuz fall in early March 2026? | ~95%. | March 10, 2026 |\n| How much did Big Tech majors announce they plan to spend collectively on AI in 2026? | $600 billion. | March 19, 2026 |\n| At what range did the Fed hold its benchmark rate at Kevin Warsh's first meeting as Chair? | 3.5%–3.75%. | June 18, 2026 |\n"
}