AI R&D Digest

AI R&D Digest is a public agent skill and feed pipeline for AI industry updates. It turns centrally generated daily archives into a concise digest of builder posts, official AI product/research blogs, and podcast episodes.

The project has two audiences:

• Subscribers install the packaged skill and ask their agent for a digest.
• Maintainers generate the public archives, update the packaged skill, and keep the source feed healthy.

Subscribers do not need Folo, X/Twitter, RSSHub, podcast transcript, or source API keys. The source archive is generated centrally and published as JSON.

Quick Start for Subscribers

You can send either of these directly to your agent:

English: Install the AI R&D Digest skill from https://raw.githubusercontent.com/wadjj/AI-RnD-Digest/main/dist/ai-rnd-digest.skill , then run /follow-airnd to set it up.

中文：请安装 AI R&D Digest skill： https://raw.githubusercontent.com/wadjj/AI-RnD-Digest/main/dist/ai-rnd-digest.skill ，然后运行 /follow-airnd 完成设置。

1. Install the skill from https://raw.githubusercontent.com/wadjj/AI-RnD-Digest/main/dist/ai-rnd-digest.skill.
2. Say set up AI R&D Digest or invoke /follow-airnd.
3. Choose digest frequency, delivery time, language, and delivery method.

Defaults:

• Digest window and delivery interval: 2 days
• Delivery time: 09:00 in the agent/platform default timezone
• Language: Chinese
• Delivery method: in-chat/stdout unless the user explicitly selects a supported external channel

For external delivery channels, the final digest should be generated once, written to /tmp/ai-rnd-digest.txt, then delivered by reading that exact file. Delivery must not retype, shorten, compress, or regenerate the digest unless the user explicitly asks for a compressed version.

After setup, run /follow-airnd whenever you want the latest digest.

What the Digest Contains

The digest is built from the checked-in source configuration and generated public archives:

• X / Twitter posts from AI builders and operators
• Official blogs from AI companies, research teams, product teams, and technical writers
• Podcast episodes when available in the public archive

Private inbox items are not part of the public subscriber archive.

Output Format

Final digests are optimized for phone reading and visible source attribution. The digest starts with one H1 title, keeps H2 section headers, and each person or content item starts with a bold label followed by the summary body and raw source URLs as bullets:

# AI R&D Digest - 2026-06-11

## X / TWITTER

**Full Name, role/company**

Summary text.

Sources:
- https://example.com/source-1
- https://example.com/source-2

URLs must stay visible as raw text. The skill should not hide links behind labels such as [Original](url) or [原文](url).

Local Commands for Maintainers

npm run generate
npm run generate:force
npm run generate:preview
npm run prepare
npm run package:skill
npm run deliver
npm run sync:folo-token -- --dry-run
npm test

Command summary:

• npm run generate fetches the central Folo list and writes the daily archive.
• npm run generate:preview ignores existing state and does not update state-feed.json.
• npm run generate:force allows a same-day rerun and merges into the existing archive.
• npm run prepare emits the JSON blob an agent needs to remix the digest.
• npm run package:skill rebuilds dist/ai-rnd-digest/ and dist/ai-rnd-digest.skill from the top-level skill, prompts, scripts, and config.
• npm run deliver sends final digest text according to subscriber delivery config. Digest runs should pass --file /tmp/ai-rnd-digest.txt so the delivered content is identical to the generated artifact.
• npm run sync:folo-token syncs a local Folo CLI session token to GitHub Actions.

Generated Files

• archives/YYYY-MM-DD.json: one daily incremental archive with X, blog, and podcast sections
• feed-index.json: retained archive list and per-archive stats
• state-feed.json: shared dedupe state
• dist/ai-rnd-digest.skill: packaged subscriber skill
• dist/ai-rnd-digest/: unpacked packaged skill contents

Run npm run package:skill after changing SKILL.md, prompts/, scripts/prepare-digest.js, scripts/deliver.js, or config/default-sources.json.

Generation is incremental. The daily job uses short source lookbacks as a miss-recovery window, but only items not already present in state-feed.json enter the day's archive. Subscriber frequency does not change source fetching; it only chooses how many daily archives to merge for the final digest.

Maintainer Generation

GitHub Actions is the preferred generation path because workflow concurrency serializes scheduled runs.

Repository setup:

1. Set repository secret FOLO_TOKEN.
2. Optionally set repository variable FOLO_LIST_ID; otherwise the checked-in config is used.
3. Optionally set repository secret AITRENDPUSH_PRIVATE_RSS_FEEDS for private paid RSS feeds.
4. Let .github/workflows/generate-feed.yml run daily and commit archive updates.

AITRENDPUSH_PRIVATE_RSS_FEEDS is a JSON array. Keep it in repository secrets, not repository variables:

[
  {
    "id": "stratechery-paid",
    "name": "Stratechery",
    "type": "blog",
    "url": "https://example.com/private/rss",
    "urlStrategy": "guid",
    "contentMode": "plainText",
    "stripQuery": true
  }
]

Private RSS URLs are fetch-only inputs. The generator writes only sanitized blog items to archives/YYYY-MM-DD.json, uses canonical item URLs for state, and fails before persisting output if an access_token or configured private token appears in generated JSON.

Local generation is available for manual runs or fallback automation:

npm run generate

For local schedulers, set REPO_DIR to your checkout path instead of hard-coding a machine-specific directory:

REPO_DIR="/path/to/AI-RnD-Digest"
cd "$REPO_DIR"
npm run sync:folo-token -- --yes --min-valid-days 7 --skip-if-remote-valid --renew-if-needed --pushover-on-login-needed

Folo Token Sync

GitHub Actions needs a valid FOLO_TOKEN repository secret. Maintainers can check the local token and GitHub access without writing:

npm run sync:folo-token -- --dry-run

To write the secret after an explicit confirmation prompt:

npm run sync:folo-token

The script never prints the token. It reads the local Folo CLI session, checks its expiration with Folo, verifies GitHub access, and writes the repository secret via gh secret set.

When --skip-if-remote-valid is set, the script first checks the non-sensitive repository variable FOLO_TOKEN_METADATA. If the GitHub secret exists and the metadata says the remote token is still valid for at least --min-valid-days, the script exits without reading the local Folo token or writing the secret.

Optional emergency alert environment variables:

export PUSHOVER_APP_TOKEN="your-pushover-app-token"
export PUSHOVER_USER_KEY="your-pushover-user-key"
export PUSHOVER_DEVICE="optional-device-name"
export PUSHOVER_SOUND="siren"
export PUSHOVER_RETRY="60"
export PUSHOVER_EXPIRE="3600"

With --pushover-on-login-needed, the script sends an emergency Pushover message only when manual Folo login is actually required.

Manual refresh:

npx --yes folocli@latest login
npm run sync:folo-token

Subscriber Runtime

Reader configuration lives in the subscriber's local agent config directory as config.json and is managed by the skill during setup:

{
  "platform": "openclaw",
  "language": "zh",
  "frequencyDays": 2,
  "deliveryTime": "09:00",
  "delivery": { "method": "stdout" },
  "onboardingComplete": true
}

frequencyDays is clamped to 1-7. It controls both the digest window and the delivery interval.

Source Changes

The source list is maintained in config/default-sources.json. Source additions or removals are maintainer changes, not subscriber preferences.