Skip to content

CLAUDE.md

Latest commit

 

History

History
87 lines (68 loc) · 4.66 KB

CLAUDE.md

File metadata and controls

87 lines (68 loc) · 4.66 KB
Raw

MorningAI

Daily AI news tracking skill that runs inside AI coding tools (Claude Code, Codex, OpenCode, Hermes, Cursor, ClawHub). Not a standalone service — no Docker, no servers. Invoked via /morning-ai or by reading SKILL.md.

Tech Stack

  • Language: Python 3.9+
  • Dependencies: stdlib only — no requirements.txt, no pyproject.toml. Uses urllib, json, xml.etree, smtplib, email.mime, string.Template, dataclasses.
  • Distribution: source files only. Plugin manifests for Claude Code (.claude-plugin/) and Codex (.codex-plugin/).
  • Schedule frequency: daily at 08:00 UTC+8 (idempotent, unattended).

Project Structure

SKILL.md                       # Main entry point (symlinked from skills/morning-ai/)
AGENTS.md                      # Auto-discovery for OpenCode/Hermes/Codex
.claude-plugin/                # Claude Code plugin metadata + marketplace
.codex-plugin/                 # OpenAI Codex plugin metadata
lib/                           # Shared Python modules (collectors, scoring, dedupe, IO)
skills/
├── morning-ai/                # Main orchestrator (SKILL.md is a symlink to root)
├── tracking-list/             # Step 1: Python collectors → data_{date}.json
├── leaderboard/               # Optional: leaderboard ranking diff over time
├── gen-infographic/           # Step 4: image generation (5 styles)
├── gen-social/                # Step 5: per-platform social copy (X, Xiaohongshu)
├── gen-message/               # Step 6: message digest for chat platforms
└── gen-email/                 # Step 7: SMTP email subscription
entities/                      # 80+ tracked entities organized by group (markdown tables)
templates/                     # Report templates
docs/                          # Style previews, email setup guide, etc.
tests/                         # Unit tests (stdlib unittest)

Pipeline

All sub-skills consume the same data_{YYYY-MM-DD}.json produced by Step 1, so each output channel sees identical scored / deduped / verified items.

collect.py (Step 1) → data_{date}.json
  → tracking-list/SKILL.md (Step 2: spec for AI agent)
  → report_{date}.md (Step 3: structured report)
  → gen-infographic/ (Step 4, optional: cover + per-type images)
  → gen-social/ (Step 5, optional: per-platform copy)
  → gen-message/ (Step 6, optional: chat digest)
  → gen-email/ (Step 7, optional: SMTP delivery)

Configuration

Loaded by lib/env.py in this priority order (later overrides earlier):

  1. ~/.config/morning-ai/.env (global)
  2. .claude/morning-ai.env (project)
  3. .env (project, gitignored)
  4. .env.local (local-only)
  5. Environment variables matching prefixes: MORNING_AI_, GITHUB_, IMAGE_GEN_, IMAGE_STYLE, GEMINI_, MINIMAX_, SOCIAL_, MESSAGE_, EMAIL_

When extending with a new feature, add its env-var prefix to the whitelist tuple in lib/env.py:get_config().

Output Files (date-stamped, gitignored)

Pattern Producer Notes
data_{date}.json tracking-list/collect.py Source of truth for all downstream skills
report_{date}.md morning-ai (Step 3) Detailed daily report
manifest_{date}.json gen-infographic Image generation jobs
social/{date}_*.md gen-social Per-platform copy
message_{date}.md / .png gen-message Chat digest
email_{date}.html / .txt gen-email Local previews of sent emails
email_{date}_manifest.json gen-email Per-recipient send status

Key Conventions

  • Skill versioning: bump every version: "x.y.z" field in SKILL.mds + plugin.jsons + marketplace.json together. Add a CHANGELOG entry.
  • Symlink: skills/morning-ai/SKILL.md is a symlink to root SKILL.md so Claude Code's skill scanner finds it. Edit the root file — both paths show the same content.
  • Date convention: YYYY-MM-DD in UTC+8 (Asia/Shanghai). The 24-hour collection window is 08:00 UTC+8 yesterday → 08:00 UTC+8 today.
  • Sub-skill activation: each optional output channel is gated by an *_ENABLED=true env var (MESSAGE_ENABLED, SOCIAL_ENABLED, EMAIL_ENABLED). Steps silently no-op when their gate is unset.
  • Script entry pattern: scripts under skills/*/scripts/ add sys.path.insert(0, str(Path(__file__).resolve().parents[3])) to import lib.
  • Stdlib only: do not introduce external dependencies. The project's portability across AI tool environments depends on this — anything more than python3 is friction.

Testing

python3 -m unittest discover tests -v

For end-to-end checks of an output channel, prefer dry-run modes (e.g. EMAIL_DRY_RUN=true) over hitting real services.