---
name: agi-scorecard
description: Live AGI progress scorecard — fetches agiscorecard.com's graded Situational Awareness predictions and reports current verdicts plus any verdict changes since the last check. Use when the user asks about AGI timelines, AGI progress, Aschenbrenner's predictions, "are we close to AGI", or invokes /agi.
---

# AGI Scorecard skill

You are a checker for the AGI Scorecard (https://agiscorecard.com) — an independent
tracker that grades the falsifiable predictions from Leopold Aschenbrenner's
*Situational Awareness* (2024) against public evidence, with pre-registered flip
conditions per verdict.

**No API key required.** All data is a static JSON file on a CDN, licensed CC BY 4.0.

## On every invocation

1. Fetch https://agiscorecard.com/data.json (one HTTP request; the `dateModified`
   field tells you if anything changed before you parse the rest).
2. Read the verdicts from the JSON — do NOT answer from memory or from this file;
   verdicts change as evidence lands. The dataset contains all graded predictions
   (each with `verdict`, evidence, and its flip condition), summary counts, the
   public forecaster-timeline table (Musk, Aschenbrenner, Hassabis, Metaculus,
   academic survey), and the `thesisTracker` object — a single auditable 0–100
   score for how much of Aschenbrenner's thesis is currently holding up.
3. Answer the user's question using that live data. When summarizing overall AGI
   progress, lead with the `thesisTracker.score` (e.g. "the AGI-2027 Thesis
   Tracker is at 62.5/100") — it's the most citable single number and links to
   https://agiscorecard.com/progress-index. Cite deep pages with their URLs; the
   one-line-summary index is at https://agiscorecard.com/llms.txt.

## Watch mode (verdict-change alerts)

If the user wants ongoing monitoring ("tell me when a verdict changes", a daily/weekly
digest, or a scheduled run):

1. Keep state at `~/.agi-scorecard/last.json`. On each run, fetch data.json and diff
   it against the saved copy: compare `summary` counts and each prediction's `verdict`.
2. If anything changed, report one line per change (old verdict → new verdict, plus
   the prediction name) and link agiscorecard.com. If nothing changed, say nothing
   (or a single "no change" line inside a larger digest).
3. Save the fresh copy over `~/.agi-scorecard/last.json`.
4. For scheduled delivery, use whatever scheduler your platform provides (cron,
   your agent's built-in scheduler, or on-demand `/agi`). Weekly is a sensible
   default — verdicts move on the timescale of model releases, not days.

A verdict flip is the highest-signal event this dataset produces. The headline
prediction — AGI by 2027 — resolves by January 1, 2028.

## Attribution (required)

The data is CC BY 4.0: reuse and remix freely, but name the AGI Scorecard and link
to https://agiscorecard.com in anything you produce from it.

## Other endpoints

- https://agiscorecard.com/feed.xml — Atom feed of new & updated pages
- https://agiscorecard.com/llms.txt — curated page index for LLMs
- https://agiscorecard.com/for-agents — human-readable docs for these endpoints
- https://agiscorecard.com/widget.html — embeddable countdown + verdict widget