agent_skills

Calibrate the Claude Code statusline against the real /usage numbers. Use when the user shares a screenshot of the in-app `/usage` panel (or the claude.ai Settings → Usage page) and wants the statusline's "time left", session %, and weekly % to match what Claude actually reports. Fixes ccusage block-drift, where the local 5-hour clock starts from the first transcript message instead of the real server-side reset.

Dispatch a review task to Gemini CLI (large-context review — plans, docs, large file corpora) or Codex CLI (code review — diffs, PRs, implementation quality). Routes based on the nature of the artifact being reviewed. Use from achieve-goal, literature-survey skills, or any skill that needs a CLI-based review gate. This is a headless CLI review gate over a concrete artifact (code, diff, or doc corpus); for an interactive second opinion or a side-by-side model comparison via MCP chat, use consult-llm instead.

Use when the user wants a second opinion from Gemini or OpenAI/ChatGPT, asks to "consult" another model, or wants to compare responses across models. This is the interactive MCP chat path for a second opinion or model comparison; to run a structured review gate over code, diffs, plans, or large doc corpora through a headless CLI, use cli-review instead.

Use this skill when deciding whether and how to update the knowledge base at /path/to/works/for/you/knowledge_base/ after completing a task. This is the post-task "does this work belong in the KB, and where" decision path; to turn a pasted LLM conversation transcript into a KB note use kb-from-conversation instead.

Periodic review of the memory system at ~/.claude/memory/ for duplicates, redundancies, conflicts, and stale entries. Use when the user asks to audit, clean up, or review memory. Presents findings with enforcement trade-off context — never retires anything without user confirmation. Scoped to the agent memory files under ~/.claude/memory/; for the same kind of duplicate/stale/conflict audit over the knowledge base at /path/to/works/for/you/knowledge_base/ use kb-sweep instead.

Use this skill when the user wants to "promote", "move", "upgrade", or "migrate" a project-level skill, agent, or rule (e.g. a Cursor rule, a project-local AI rule, or any .md skill/agent file) into a global Claude scope. By default always promotes to global (~/.claude/skills/, ~/.claude/agents/, ~/.claude/CLAUDE.md) unless the user explicitly says "project-level". Moves the file so it is shared across all projects and symlinks it back so the original project is not broken. This keeps the skill private to your global ~/.claude scope; to publish skills to the public agent_skills GitHub repo, use publish-skills instead.

Security-reviewed git commit for the ~/.claude backup repo — scans staged changes for secrets before committing. Use when committing only the ~/.claude config repo on its own; for the multi-repo "save state" checkpoint that scans and commits the knowledge base, paper-reading, agent_skills, and ~/.claude repos together, use save-to-github instead.

Surveys informal and industry sources for a given topic — covering Hacker News, engineering blogs, ProductHunt, Medium, Substack, X (indirect), and Quora. Complements literature-survey (which targets academic papers) with community signal, practitioner discourse, and tooling trends. Use when the user wants a community survey, practitioner signal, tooling trends, or runs /community-survey <topic>. This is the one-shot variant; to track a topic's community signal over time (trajectory queries — what shifted, went quiet, or is new across runs) use community-survey-nlm instead.

Use when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies. This is the methodology for grouping independent problems into focused parallel agent briefs; for deterministic scripted fan-out use the built-in Workflow tool, and for a single subagent use the Agent tool directly.

Literature Survey Skill - performs a structured literature survey for a given topic and saves outputs under the paper reading repo's literature-survey/[topic]/ folder. Invoke with /literature-survey <topic> [topic-slug]. Use this lightweight one-shot variant for a single survey; for ~95% token savings and persistent topic memory across runs (cross-run trajectory queries over a NotebookLM notebook) use literature-survey-nlm instead.

Analyzes ML/AI research papers from PDF files or URLs. Summarizes papers, critiques experiment design and statistical validity, assesses industry contribution, compares to prior work via web search, and checks dataset availability. Use when the user asks to read, analyze, or review a paper, provides a PDF path or arxiv/paper URL. Do NOT use to review the user's own unpublished draft before submission — use paper-submission-review for that.

Reviews a paper draft (PDF, LaTeX source, or extracted text) before submission and returns prioritized editing suggestions with rationales. Use when the user asks to "review a paper draft", "check a paper before submission", "do a pre-submission review", or supplies a paper file with the goal of polishing it for a venue. Findings cover claim–evidence calibration, contribution naming, abstract/intro framing, section structure, operational clarity, reproducibility anchors, metric choice, baseline credibility, table/figure design, result interpretation, citation discipline, notation hygiene, cleanup, conclusion alignment, and length discipline.

Use when the user wants help reasoning through a problem, presents an argument or plan and asks for feedback, says "help me think about X", "poke holes in this", "is this a good idea?", or otherwise invites scrutiny of their reasoning. Applies first-principles decomposition and critical-thinking checks to evaluate and strengthen the user's argument or thought process.

Generic skill for tasks that require planning, debugging, or iteration before they can be completed. Three modes — execute / optimization / orchestrate — selected by the request shape. Discriminator: does the task require planning? If the request is unambiguous and mechanical ("run this notebook with python3"), do NOT use this skill — direct execution is correct. If achieving the goal requires planning, recovery from failures, or iterating toward a target, this skill applies. Triggers: "use achieve-goal to X", "iterate to achieve metric Y", "execute this multi-milestone plan", or any goal that needs planning + recovery.

Generates a new Jupyter notebook with three Colab bootstrap cells: Drive mount, work directory under Drive, and GitHub clone + branch checkout. Pure scaffolding — does not execute the notebook. Triggers: "create a Colab notebook", "bootstrap Colab setup", "new Colab notebook with Drive mount and clone", "scaffold a Colab notebook for this repo".

Community survey skill backed by a persistent NotebookLM notebook per topic. Same source platforms as community-survey, plus cross-run trajectory queries (how has X shifted, what's gone quiet, what's new since N months ago). Use when the user wants to track a topic's community signal over time.

Review cowork project outputs and sync the workspace-level README.md and log.md in cowork_projects/. Use when cowork agents have done work and the workspace index needs to be updated, or when the user says "sync cowork", "review cowork status", or "update cowork workspace".

Use when the user reports a bug, asks to debug specific behavior, or when an unexpected failure occurs during execution. Applies the standing debugging protocol — reproduce, minimize, validate fix on a single job before scaling.

Resolves paper titles and repo-relative PDF paths from the curated Awesome-Deep-Learning-Papers-for-Search-Recommendation-Advertising index by keyword. Returns section, title, decoded path inside the companion PDF repo, and the GitHub blob URL. Triggers: user names a paper / acronym / author / venue and wants the canonical path or URL from the Awesome rec-sys list (e.g. "find DIEN", "where is the wide & deep paper", "Awesome rec-sys list").

Looks up the local path of a named repository by reading the repos registry at /path/to/works/for/you/knowledge_base/context/registry/repos.md. Use this whenever you need a repo's path and don't already know it — do not guess or hardcode paths.

Looks up which Claude skills and kb/context docs apply to a named repository and returns the paths. User-invoked at the start of a repo session. Companion to fetch-repo-path. Use when the user says "/fetch-repo-skill <repo>", "load skills for <repo>", or "what skills apply to <repo>?".

Reviews deep learning code (.py files and Jupyter notebooks) to ensure GPU is used when available. Identifies anti-patterns like missing device detection, hardcoded CPU usage, tensors not on GPU, and performance issues. Supports PyTorch, TensorFlow/Keras, and JAX. Triggers: "review code for GPU usage", "check GPU utilization", "audit device placement", "ensure DL training runs on GPU".

Orchestrates a knowledge base housekeeping sweep: checks KB size and git log to scope the work, then delegates to one or more knowledge-base-house-keeper agents. Use when the user wants to audit the KB for duplicates, conflicts, stale content, or broken references.

NotebookLM-backed literature survey skill. Token-efficient alternative to literature-survey: uses NotebookLM notebook_query for per-paper analysis (~95% token savings). Invoke with /literature-survey-nlm <topic> [topic-slug].

Runs a Jupyter notebook on Google Colab via SSH + papermill. Uses scp for fast file sync (no git round-trip), handles remote execution, and surfaces errors for fix-and-retry loops. Triggers: "run notebook on Colab", "run notebook with Colab", "execute notebook on Colab", "test notebook on Colab".

Use this skill when the user says "save the state to github", "save state", "push state to github", or "checkpoint to github". Performs a security risk scan, ignores risky files, stages safe ones, commits, and offers to push. Operates on four repos in every run: the knowledge base repo at /path/to/works/for/you/knowledge_base/, the Claude config repo at ~/.claude/, the paper reading repo at /path/to/works/for/you/Projects/paper_reading_repo/, and the public agent skills repo at ~/agent_skills/.

Records unexpected agent behavior or failures to a persistent ledger at building_knowledge_and_agent/agents-failure.md. Use when the user reports that an agent, subagent, skill, hook, or the lead agent itself behaved unexpectedly, ignored a rule, produced wrong output, or skipped a documented step. Groups similar failures by a loose fingerprint and prompts the user to fix the root cause once the same failure type has been observed three times.

Creates a dated ad-hoc project folder under knowledge_base/projects/ad_hoc/ named YYYYMMDD_{topic}, sets it as the session working directory, and tells the agent to save all session artifacts there. Use when the user says "create an ad hoc project for X" or "start an ad hoc project on X".

Converts a raw LLM conversation transcript into a structured knowledge base note and saves it in the correct project folder under `/path/to/works/for/you/knowledge_base/projects/`. Use when the user pastes an LLM conversation and asks to "save this", "document this", "add to the KB", or "jot this down", or when a conversation contains reusable frameworks, plans, or decisions worth retaining.

Use this skill when working with Google NotebookLM — creating/querying notebooks, managing sources, or running research workflows via the notebooklm-mcp MCP server.