OpenClaw秘書OS best practice intake — 2026-05-31

OpenClaw公式docsと現代agent運用best practiceを、ひめのの運用OSへ落としたメモ。

結論

OpenClaw/ひめの改善は、自律性を盛るより先に『小さい巡回レーン』『決定的スクリプト』『証拠付きdone』『View+ソース』で固めるのが今のbest practice。でかい自動改造AIにしない。ここ大事。

巡回: heartbeat is a patrol loop, not an error-only monitor: rotate Discord/docs/feedback/report/task/runtime lanes.
反映: every useful signal must land in one of memory, task ledger, TOOLS/AGENTS/SOUL, skills, or View.
証拠: done requires evidence: command output, test, lint, View URL, source link, or named blocker.
View: human-readable research/proposals must become View pages, not just Markdown dumps.
ソース: external-source claims need source links in the report/View body.
小さく: prefer small deterministic scripts and focused lanes over one giant autonomous blob.

heartbeat_active_search に docs/best-practice intake, report quality audit, Discord OpenClaw intel をレーンとして持たせる。
レポート生成は scripts/view_report.py を通し、View URLを返す。
Discord intel は raw log ではなく、テーマ・反映候補・ソース付きViewへ圧縮する。
done は scripts/task_ledger.py lint と対象スクリプトのテスト/実行結果を付ける。

OpenClaw Agent workspace: https://docs.openclaw.ai/concepts/agent-workspace — workspace is private agent home; keep config/credentials/sessions outside git
OpenClaw Automation & tasks: https://docs.openclaw.ai/automation — cron for exact timing, heartbeat for contextual patrol, tasks for detached work
OpenClaw Heartbeat: https://docs.openclaw.ai/gateway/heartbeat — heartbeat is periodic main-session turn; use HEARTBEAT.md as tiny checklist
OpenClaw Memory: https://docs.openclaw.ai/concepts/memory — plain Markdown memory; only written facts survive
OpenClaw Skills: https://docs.openclaw.ai/tools/skills — repeatable procedures should become workspace skills
OpenClaw Sub-agents: https://docs.openclaw.ai/tools/subagents — long/parallel work should be isolated and push-based
Anthropic: Building effective agents: https://www.anthropic.com/research/building-effective-agents — prefer simple composable workflows; use agents when flexibility is worth cost
Hamel Husain: LLM Evals FAQ: https://hamel.dev/blog/posts/evals-faq/ — use real traces, error analysis, and iterative eval refinement
HumanLayer: 12 Factor Agents: https://www.humanlayer.dev/blog/12-factor-agents — own prompts/context/control flow; tools are structured outputs into deterministic code