daily-research
Autonomous daily research powered by claude -p (Claude Code's non-interactive mode) — each morning the repo lines of a rotation window ([general] lines_per_day) each run call 1 (Opus research) inside the line's repo (cwd = the repo) so it reads the repo's own operational context, picks one external development relevant to the repo as a question, explains what happened, why it matters, and how the repo can use it, and writes a free-form explanatory note; call 2 (Sonnet by default, fresh context, fail-open) then revises the note for first-contact readability and re-fetches every claim to check it against its source (ADR-0010 / ADR-0014 / ADR-0017). A run may also end as a short brief, or honestly as quiet (nothing new) or hold — the state it declares in its run receipt is what the report gate reads. Notes go — readable with zero prior context — directly to an Obsidian vault. Deadline-bound opportunities are captured in a fixed 機会メモ tail section; notes carry no proposals or approval requests (ADR-0009).
Skill anatomy: prompts/repo-research-protocol.md is the skill's core intelligence; prompts/ecosystem-research-protocol.md is its sister for lines declared mode = "ecosystem", which have no repo and run from their own state/<line>/ directory (same receipt, gate, content states, and tail sections — ADR-0018). prompts/verification-protocol.md drives call 2 (clarity first, then source verification). scripts/daily-research.sh is a thin wrapper that picks the day's lines, invokes claude -p twice per line with those protocols, and decides each line's success from its run receipt. Everything else (tests, launchd plists, config template) is supporting infrastructure.
When to use
- Scheduled daily execution — macOS
launchd(included) or any cron / systemd setup - Manual one-shot run —
./scripts/daily-research.shor invoke via/daily-researchafter installing as a Claude Code skill
Design philosophy
One question, answered honestly. A run lists 3–6 candidate questions and keeps the one whose answer is least already known, then explains it for a reader with zero prior context, keeping the source's claim, the inference, and the implication apart (ADR-0014 / ADR-0017). A morning with no material ends as a brief or a plain "nothing new" rather than a padded article. The pipeline's capacity is defined by human attention, not by source count — hence a fixed-width rotation window — and the filters (question screen, citation gate, and a separate process that re-fetches every claim) rather than downstream storage are where quality is enforced. Every claim carries an as-of date and every deadline-bound opportunity an expiry date: knowledge here goes stale on a one-week scale.
Execution
./scripts/daily-research.sh
Prerequisites, configuration, and scheduling are documented in the main README. Operations details (monitoring, troubleshooting) are in RUNBOOK.
Install as a Claude Code skill
git clone https://github.com/shimo4228/daily-research.git \
~/.claude/skills/daily-research
After cloning, Claude Code recognizes SKILL.md and the skill becomes invocable as /daily-research. For automatic daily execution, follow the launchd or cron setup in the main README.
Documentation
- README.md / README.ja.md — overview, features, quick start, design decisions
- docs/RUNBOOK.md / docs/RUNBOOK.ja.md — operations guide
- docs/CONTRIB.md / docs/CONTRIB.ja.md — development guide
- config.example.toml — configuration template