feed-digest

A shareable agent skill that produces a daily digest from your feeds — Twitter (via community archive), Substacks, blogs. Memory-aware. Self-improving: you give feedback in your 1:1 agent chat, next day's digest adapts.

Built by Francisco (Xiq) and Ivan, intended to be portable across agent harnesses (Hermes, Claude Code, anything with file-edit + Telegram/email tool access).

Status

Pipeline implemented and working. Ingest (3 adapters), digest (OpenRouter), deliver, promote, publish all functional. 28 posts ingested, first digest generated.

If you're an AI agent picking this up: read AGENTS.md first. It tells you to study gstack as a reference for skill-authoring style, explains the build order, lists the critical rules, and sets up the handoff protocol. Then docs/design.md → docs/test-plan.md for the full context.

Install

git clone https://github.com/TheExGenesis/feed-digest
cd feed-digest
./install

The install script detects whether Hermes or Claude Code is installed and symlinks the repo into the appropriate skills directory (~/.hermes/skills/feed-digest or ~/.claude/skills/feed-digest). Re-running ./install is idempotent.

How it works (once built)

Cron fires daily at your configured time (default 07:00). The pipeline:

1. Ingest — fetch new items from each configured source into feeds/raw/<type>/<source-id>/<date>.md.
2. Compose — read past 7 days of digests + your memory folder + accumulated feedback, produce digests/YYYY-MM-DD.md.
3. Deliver — your host agent reads the new digest and sends it to you via email and/or 1:1 Telegram.
4. Feedback — you reply to the digest in your normal agent chat. Agent detects feedback and writes it to feedback/<digest-id>.md. Tomorrow's digest adapts.
5. Publish (Layer B, optional) — agent asks "publish a curated version?" You say yes. Agent runs hermes-feed publish <digest-id>. A stable-URL "link post" is generated and shared with your followers or group chat.

No listener daemon. No bot registration (the skill uses your existing agent's Telegram access). Just cron + markdown + the host agent as the orchestrator.

Config

First run copies config.example.yaml to config.yaml (which is gitignored). Edit config.yaml:

memory_path: ~/hermes/memory          # your personal memory folder (read-only)
timezone: America/New_York
email:
  smtp_host: smtp.gmail.com
  smtp_port: 587
  smtp_user: you@gmail.com
  smtp_password_env: FEED_SMTP_PASSWORD
  to: you@example.com
sources:
  substack:
    - https://xiqo.substack.com
  twitter_community_archive:
    - frsc
  blog_rss:
    - https://example.com/feed.xml

Secrets (SMTP password, any API keys) are read from env vars, never committed.

Docs

• Design — problem, premises, architecture, build order.
• Test Plan — coverage diagram, fixtures, eval scenarios.