ai-digest

Produce a consolidated AI + markets digest in two clearly-separated halves, so cutting-edge builder signal never gets diluted by business/markets coverage and vice versa. Render a clear visual divider between Part A and Part B. A longer digest is fine — favor substance and real coverage over brevity.

Customize this file. The watchlists, categories, and GitHub-scout targets below are a starting template. Edit them to your interests before your first run — that's the whole point. Anything in [brackets] is a placeholder to replace.

Freshness pass (run it — it's the best signal, costs pennies): Grok is the only frontier model grounded on live X posts. Run scripts/grok-x-pass.py twice, once per Part, with the gap-in-days since your last digest:

• Part A: python3 scripts/grok-x-pass.py --days <gap> --prompt "What are the top AI-native builders, AI-lab leaders, and tech/VC figures shipping, launching, or debating on X right now — agents, models, applied AI, notable tech takes? Attribute each to a person."
• Part B: python3 scripts/grok-x-pass.py --days <gap> --prompt "What are notable people saying on X about management consulting, venture & PE deal flow, M&A and capital markets, macro/investment trends, and forward-deployed-engineer (FDE) roles? Attribute each to a person."
  Fold each pass into its Part (one-liners + source URLs). Exit codes: 0 ok; 2 = no XAI_API_KEY (fall back to web search silently); 3 = API error (note briefly, fall back). No key? The digest still works on web search alone — the Grok pass just makes it sharper.

Pre-filter before writing anything up (cost discipline): as candidates surface from the Grok passes, web search, and HN/GitHub/HF sweeps, score each one before spending output tokens on a full write-up:

• Recency (0-2): genuinely new since the last digest vs. already-stale.
• Relevance (0-2): matches your actual stated interests (edit the rubric weights to match what you track) vs. generic AI news.
• Substance (0-1): real signal (a launch, a debate, a number) vs. engagement-bait or a repost of something already covered.
  Only write a full item for candidates scoring 3+/5. Silently drop the rest — don't narrate what was cut. This keeps the expensive step (synthesis) spent only on things that survive scoring, not everything encountered — the same discipline behind any good filter-then-generate pipeline, just applied at the instruction level rather than in code.

PART A — Frontier & Builders

1. Model Releases & Major Updates — new model releases, version updates, or notable capability announcements from the frontier labs (OpenAI, Anthropic, Google DeepMind, Meta, xAI, Mistral, Moonshot, etc.), covering the days since the prior digest.

2. What AI-Native Builders Are Doing — the most interesting things technical/AI-native people are building, shipping, or debating around agents and applied AI. Use the Part-A Grok pass, plus X-visible discourse via web search (e.g. site:x.com AI agents) and tech blogs/newsletters. Also scan Hacker News (news.ycombinator.com front page + "Show HN") — high-signal builder/engineer discourse that often surfaces things before or differently than X; pull 2-4 genuinely notable AI/agent items, skip generic tech.

   Watchlist (your substitute for living on X): each run, surface what these accounts have posted/shipped. Rotate through it — only surface accounts with something genuinely notable this cycle. Edit this list to match who you actually want to track.

   • AI/agent builders & researchers: Andrej Karpathy, Andrew Ng, Jim Fan, Simon Willison, swyx, Harrison Chase, Yann LeCun
   • Lab & frontier-company leaders: Sam Altman, Greg Brockman, Demis Hassabis, Aravind Srinivas, Clément Delangue (Hugging Face), Arthur Mensch (Mistral), Logan Kilpatrick (Google AI Studio)
   • Broader tech / VC leaders: Paul Graham, Garry Tan (YC), Naval Ravikant, Marc Andreessen, Satya Nadella
   • Curators (catch what the above amplify): Rowan Cheung (The Rundown AI), DAIR.AI (paper threads)
     Attribute each find to the person with a source link.

3. GitHub Signal (personalized) — a personalized repo scout, not a generic trending list.

   • (a) Your taste profile: check [your GitHub profile URL] (starred + following tabs) for anything new; note drift.
   • (b) Trending, relevance-ranked: sweep github.com/trending (daily + weekly) plus 2-3 targeted searches across [your interest areas — e.g. AI agents / MCP servers / finance tooling / PKM]. Also scan Hugging Face trending (huggingface.co/models?sort=trending) for notable open-weight model drops. Rank by relevance to you, not raw stars; one line of "why it's worth your look" per pick tied to your interests. 4-7 picks max; be skeptical of inflated trending entries.

4. Other Notable AI News — AI funding rounds, major partnerships, notable AI regulatory news.

PART B — Markets, Deals & Careers

Intentionally broader than AI — make sure the non-AI categories get real representation. Source from the open web (firm publications, deal news, reputable press, VC/consulting newsletters) + the Part-B Grok pass. Group loosely by theme:

• Consulting & Strategy — McKinsey / Bain / BCG and top firms: notable reports, insights, moves (esp. AI or market-trend related).
• Macro & Investment Trends — broad market/macro themes, not just AI.
• Capital Markets & M&A — notable large deals, PE/VC mega-rounds, M&A across sectors.
• Careers — notable role/company moves and discourse in the tracks you care about (e.g. FDE / product / BD / strategy).
• The Established AI Stack — infra, chips, model providers, enterprise application cos — the funded/established layers (leave the bleeding-edge builder layer to Part A).

Roughly 8-14 items across these; author/source + one-line insight + link for each. Skip pure self-promotion and content-free engagement-bait.

Keep each item a one-line takeaway with a source link. If something is unusually significant, end the whole digest with one line on why it matters. End your output with the full formatted digest as your final message.