agent-utils

Manage a bookmark research pipeline — fetch new bookmarks, triage against research interests, spawn parallel deep-dive agents, and review results. Use when the user says "bookmark sync", "check bookmarks", "review bookmarks", "triage bookmarks", "deep dive bookmarks", "run bookmark pipeline", or asks about setting up the bookmark system.

Update project context files when reality has drifted from docs. Use when a milestone is hit, the plan has changed, CLAUDE.md feels stale, or the user says "update the docs", "sync context", "evolve context". Walks through ledger, plan, CLAUDE.md, README, and architecture docs.

Create a new project with consistent structure and documentation. Use when the user says "new project" or "start a project" or when building something new. Sets up for agent-assisted development from day one.

Turn any question into an automated feedback loop. The pattern — design a sensor (eval), an actuator (code the agent modifies), and launch an evolutionary optimization loop. Works on any (code → run → score) problem — model training, software optimization, prompt tuning, eval design, data pipelines, agent harnesses. The eval function IS the research.

Mine session logs for skill usage patterns, corrections, and failures to propose improvements. Use when the user says "improve skills", "review skill usage", "what have we learned about our skills", or periodically to evolve skills based on real usage data.

Design and improve harnesses — the environments, constraints, feedback loops, and control systems that make AI agents do reliable work. Use when setting up a new project for agent-assisted development, diagnosing why agents are producing bad output, reviewing architectural constraints, or thinking about how to scale human oversight of agent workflows. Synthesizes OpenAI's harness engineering principles with cybernetic control theory.

Structured literature review workflow. Use when the user asks to "do a lit review", "review this paper", "triage these papers", "compare papers", or mentions arXiv, research papers, or academic literature.

Write and maintain research logs — durable records of experiments, sessions, and findings. Covers both internal experiment logs (what we ran, what happened) and published research logs (what we explored, what we learned). Use after experiments, evaluations, research sessions, or when synthesizing a day's work.

Log work to daily notes with timestamps. Use proactively after completing tasks, reaching milestones, or finishing research sessions. High signal, no redundancy.

Prepare a project for open source release. Use when the user says "prep for oss", "prepare for open source", or similar. Scans, discusses, and transforms projects for public sharing.

Commit notes and vault changes to git when substantial. Use proactively after significant work in a markdown vault (Obsidian or any markdown directory) — structural changes, new notes, reorganization. Think like merging a PR - meaningful bundles.

Persistent memory system for Claude Code. Understands the session dump + vault search + re-injection pipeline. Use when discussing memory architecture, debugging context loss, or explaining how the agent-memory plugin works.

Strategic thinking for knowledge work — orient, prioritize, plan, and stress-test. Default mode is "given everything in context, what should I work on?" Also handles daily/weekly planning, meeting prep, and plan interrogation. Socratic by default. NOT for code planning (use Claude's plan mode for that).

Write substantive research notes — comprehensive reference documents that teach something meaningful about a topic.