Automated daily research reports powered by Claude Code non-interactive mode and macOS launchd.

Zero Python. Just shell scripts + prompt files.

Every morning at 5:00 AM, Claude autonomously browses the web, selects trending topics, conducts multi-stage deep research, and writes reports directly into your Obsidian vault.

launchd (AM 5:00)
  └─ daily-research.sh
       ├── MCP health check (Haiku)
       │     ├── OK → Mem0 enabled
       │     └── Fail → continue without Mem0
       │
       ├── Pass 1: Opus (theme selection)
       │     ├── Read config.toml       # What to research
       │     ├── Read past_topics.json  # Avoid duplicates
       │     ├── WebSearch              # Latest trends
       │     └── Score & select 2 topics
       │
       ├── Pass 2: Sonnet (research & writing)
       │     ├── Mem0 search-memories   # Recall past insights (if available)
       │     ├── WebSearch x 20-30      # Multi-stage research
       │     ├── WebFetch (primary sources)
       │     ├── Write 2 reports        # → Obsidian vault
       │     └── Mem0 add-memory        # Store key findings (if available)
       │
       └── Eval: Opus (quality scoring, non-fatal)
             ├── 6 dimensions x 2 reports
             └── Append to scores.jsonl

2-pass architecture: Opus handles theme selection (deep reasoning), Sonnet handles research and writing (speed + cost efficiency). If Pass 1 fails, Sonnet handles everything as a fallback.

The key insight: Claude Code's -p flag turns it into a fully autonomous research agent. No API plumbing, no Python, no orchestration framework. The intelligence lives in the prompt.

• 2-pass model routing -- Opus for theme selection, Sonnet for research (best of both worlds)
• Configurable research tracks -- Define topics, sources, and scoring criteria in config.toml
• Topic deduplication -- past_topics.json prevents covering the same theme twice within 30 days
• Multi-stage deep research -- Not just a summary; generates research questions, searches 20-30 times, cross-validates sources
• Weighted topic scoring -- Novelty, momentum, buildability, and "whisper trend" scores
• Obsidian-native output -- Reports with YAML frontmatter, ready for your vault
• Persistent memory (Mem0) -- Optional MCP integration recalls past research insights and stores new findings across sessions
• Robust execution -- Lock files, log rotation, auth checks, macOS notifications, automatic fallback
• Automated quality evaluation -- LLM-as-Judge scores each report on 6 dimensions (30-point scale)

# 1. Clone
git clone https://github.com/shimo4228/daily-research.git
cd daily-research

# 2. Configure
cp config.example.toml config.toml
# Edit config.toml: set vault_path and customize tracks

# 3. Make scripts executable
chmod +x scripts/daily-research.sh scripts/eval-run.sh scripts/check-auth.sh

# 4. Verify Claude auth
./scripts/check-auth.sh

# 5. Test run (manual)
./scripts/daily-research.sh

# 6. Schedule with launchd (optional)
cp com.example.daily-research.plist com.daily-research.plist
# Edit: replace YOUR_USERNAME with your macOS username
ln -sf "$(pwd)/com.daily-research.plist" ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.daily-research.plist

daily-research/
├── config.example.toml         # Research tracks, scoring, output settings
├── prompts/
│   ├── theme-selection-prompt.md  # Pass 1: Theme selection (Opus)
│   ├── task-prompt.md             # Pass 2: Research instruction (Sonnet)
│   └── research-protocol.md      # Pass 2: Research protocol (system prompt)
├── templates/
│   └── report-template.md      # Report format with YAML frontmatter
├── scripts/
│   ├── daily-research.sh       # Main entry point (launchd calls this)
│   ├── eval-run.sh             # LLM-as-Judge evaluation (post-run)
│   └── check-auth.sh           # OAuth token health check
├── evals/
│   ├── prompts/                # Judge rubrics (6 dimensions + system prompt)
│   ├── scores.jsonl            # Score log (append-only, gitignored)
│   └── scores.example.jsonl    # Schema reference
├── com.example.daily-research.plist  # launchd schedule template
├── tests/
│   ├── test-daily-research.bats     # Unit tests
│   ├── test-e2e-mock.bats          # E2E mock tests
│   ├── test-eval.bats              # Evaluation framework tests
│   └── test-log-summary.bats       # log_summary parser tests
└── docs/
    ├── RUNBOOK.md / RUNBOOK.ja.md   # Operations guide
    ├── CONTRIB.md / CONTRIB.ja.md   # Development guide
    └── plans/                       # Future expansion plans

After each successful run, an automated LLM-as-Judge evaluation scores every generated report. This runs as a non-fatal hook -- evaluation failures never block the main pipeline.

Each report is scored by Opus on 6 independent dimensions (1-5 scale, 30 points total):

Scores are appended to evals/scores.jsonl. The pipeline_version field enables before/after comparisons when the pipeline changes. See CONTRIB.md for full details.

Edit config.toml to add new tracks:

[tracks.finance]
name = "Finance & Markets"
focus = "Fintech, DeFi, and market trends"
sources = [
  "Bloomberg Technology",
  "TechCrunch Fintech",
]
scoring_criteria = [
  { name = "Novelty", weight = 30, desc = "Not covered recently" },
  { name = "Momentum", weight = 30, desc = "Actively evolving" },
  { name = "Actionability", weight = 40, desc = "Can act on this insight" },
]

The prompt files (prompts/) and report template are written in Japanese. Reports will be generated in Japanese by default. To change the output language:

1. Edit the language constraint in prompts/research-protocol.md line 10:

   - 日本語で全て出力すること  →  - Output everything in English
   

2. Translate prompts/research-protocol.md and templates/report-template.md to your preferred language.

Edit prompts/research-protocol.md to adjust:

• Number of search queries per topic
• Source validation requirements
• Report structure and length

Edit the plist StartCalendarInterval:

<key>Hour</key>
<integer>7</integer>  <!-- Change to 7 AM -->

Then reload: launchctl unload ... && launchctl load ...

• Claude Code plugins cause hangs -- If you have Claude Code plugins installed globally, their MCP servers initialize on every claude -p call, adding minutes of overhead or causing hangs. Create .claude/settings.json in the project root to disable them:

  {
    "enabledPlugins": {
      "plugin-name@marketplace": false
    }
  }

  List each of your installed plugins and set them to false. Check installed plugins with claude plugin list. There is currently no blanket "disable all" option (tracking issue).
• OAuth token expires ~4 days -- Run claude interactively periodically to refresh
• ANTHROPIC_API_KEY must be unset -- If set, Claude uses per-token billing instead of Max plan. The script handles this with unset ANTHROPIC_API_KEY
• launchd + shell profile -- launchd does NOT source .zshrc. All PATH entries must be explicit in the script and plist
• --max-turns -- Pass 1 uses 15 turns (theme selection), Pass 2 uses 40 turns (research). These are guidelines, not hard limits
• Do NOT run from inside Claude Code -- claude -p cannot be nested inside another Claude Code session

• RUNBOOK.md / RUNBOOK.ja.md -- Operations: monitoring, troubleshooting, common issues
• CONTRIB.md / CONTRIB.ja.md -- Development: testing, CLI flags, environment variables

MIT