agent-kit

Maintain the nmem knowledge base — a lint pass that hunts contradictions, duplicates, dead references, and expired entries, plus a synthesis pass that reports what changed and what is drifting. Use when the user says "lint my memory", "curate", "整理知识库", "nmem 维护", suspects memories have gone stale, or on a scheduled maintenance run.

Delegate a self-contained task to a separate streaming agent session, then summarize the result. The backend (Claude Code or Pi) is chosen automatically from the model you name — say "let opus/sonnet/claude look at it" to route to Claude Code, or "let codex/gpt look at it" to route to Pi. Use when the user wants delegation, subagents, parallel research, isolated context, resuming a delegated task, or running a task on a specific model regardless of which runtime backs it.

Multi-perspective code review using adversarial subagent debate. First gives the user a simple but detailed What/Why/How/Refs brief: what changed, why it exists, how it works, and which files, commits, or PR references support the explanation. Then spawns parallel reviewer agents (bug hunter, security auditor, architecture critic, correctness prover) that independently analyze the current branch diff, consolidates and debates findings, and produces a machine-readable review bundle plus human-readable HTML report with near-zero false positives. The report opens with a reviewer-facing overview of the PR — its purpose, what changed and why, whether the changes are necessary, and the regression and security risk — before the debated findings. Use when the user says "review", "code review", "review my changes", "check this PR", "find bugs", "audit this branch", or wants a thorough quality check before merging.

Write a decision-first implementation plan as plan.md — design decisions with alternatives and tradeoffs, phased tasks with acceptance blocks, inline review threads, and per-phase handoff notes. Use when the user wants an implementation plan, says "write a plan", "plan this out", "写个计划", wants a plan reviewed or review threads resolved, or as the Plan step of the spec-dev workflow.

Execute an existing plan.md phase by phase — catch up from the prior handoff, implement, verify the Acceptance block, review if warranted, commit, write the next handoff. Use when the user says "implement the plan", "execute the plan", "continue the plan", "next phase", "按计划实现", picks up a stalled plan in a fresh session, or as the Impl step of the spec-dev workflow.

Scout the territory before building — a map-vs-territory diagnostic plus eight lightweight techniques (blindspot pass, brainstorm prototypes, interview, references, decision-first plans, deviation notes, explainers, quizzes) for finding your unknowns before they get expensive, applied inline with no workflow ceremony. Use when the user starts unfamiliar work, says "scout this", "blindspot pass", "unknown unknowns", "我不熟这块", "帮我找盲区", can't articulate what they want, or a long-horizon task came back wrong and planning harder isn't helping.

Plan-first development workflow — the orchestrator over the lifecycle skills: grill the idea, write one plan.md, implement it phase by phase. Use when the user wants the full disciplined loop for a feature or change: "spec-dev", "plan this out and build it", "let's think this through then implement", or disciplined, reviewed code changes end to end. For a single step alone, use that step's skill directly (grill, blueprint, mason).

Expert designer skill for filesystem-backed agent runtimes. The single skill for any UI/UX work: building new interfaces or restyling/reshaping existing ones with a distinctive, intentional aesthetic direction that doesn't read as templated. Produces Preact/React prototypes, HTML artifacts, slide decks, dashboards, landing pages, mobile app screens, editorial artifacts, and polished animated UI/micro-interactions. Includes 150+ design systems, 110+ design templates from Open Design, bold frontend aesthetic guidance, deliberate typography and palette choices, motion guidance for transitions, scroll reveals, spring physics, loading skeletons, parallax, layout animation, UI polish, interface copywriting, and mandatory anti-slop rules for direct, human copy.

Remove AI writing patterns from prose and make drafts sound like a specific human wrote them. Use when drafting, editing, rewriting, reviewing, or cleaning copy, essays, docs, posts, emails, UI copy, bios, reports, or any text that feels generic, polished-but-dead, promotional, formulaic, or AI-generated. Combines comprehensive AI-writing pattern detection with a strict stop-slop final pass for direct, specific, human prose.

Stress-test an idea, design, or approach through sharp, structured interrogation. One question at a time, each with a recommended answer. Probes assumptions, edge cases, hidden dependencies, and fuzzy language until the design is either solid or abandoned. Use when the user says "grill this", "stress-test", "poke holes", "challenge this", "what am I missing", "what could go wrong", "devil's advocate", or wants their thinking pressure-tested before committing to it.

Manage files on an OpenList (AList-compatible) cloud storage server. Use when the user wants to list, search, copy, move, rename, delete, upload, or download files on cloud storage managed by OpenList/AList. Triggers on phrases like "openlist", "alist", "manage cloud storage", "network drive", or file operations against a self-hosted file server. Also use when the user mentions Emby, Jellyfin, or media library organization on cloud storage.

Create robust Python automation with full logging and safety checks. Use when tasks need complex data processing, authenticated API work, conditional file operations, or error handling beyond simple shell commands.

Improve existing code at any scale — from cleaning up a single function to restructuring an entire module hierarchy. Use this skill whenever the user wants to simplify code, refactor for readability, find architectural improvements, consolidate tightly-coupled modules, deepen shallow abstractions, reduce complexity, clean up after a feature implementation, or make a codebase more testable and navigable. Triggers on phrases like 'simplify this', 'clean up', 'refactor', 'improve architecture', 'make this cleaner', 'find refactoring opportunities', 'reduce complexity', or any request to improve code quality without changing behavior.

Lark/Feishu CLI skills for Anna sessions. Covers workspace operations — calendar, docs, tasks, mail, and messenger — via the lark-cli tool. In Anna, lark-cli runs under an env-var auth model with binaries resolved from $ANNA_HOME/bin: LARKSUITE_CLI_USER_ACCESS_TOKEN, LARKSUITE_CLI_APP_ID, and LARKSUITE_CLI_BRAND are injected at session start; no manual config init or auth login is needed. Always read lark-shared first for identity selection (--as user vs --as bot), scope / permission-denied handling, token refresh, and the Anna-specific rules that override upstream documentation.

Extract text, tables, metadata, and images from 91+ document formats (PDF, Office, images, HTML, email, archives, academic) using Kreuzberg CLI.

Send emails via the Cloudflare Email Sending REST API. Trigger this skill when the user wants to send an email using Cloudflare's email service. Use curl directly with credentials from environment variables.

Use the `gws` CLI to interact with Google Workspace services — Gmail, Drive, Sheets, Calendar, Docs, Tasks, Chat, and more. Trigger this skill whenever the user wants to read/send email, manage Drive files, update spreadsheets, check calendars, create tasks, send Chat messages, or run any Google Workspace operation. Also trigger when the user mentions Google Workspace, Google APIs, or asks to automate anything across Gmail/Drive/Sheets/Calendar/Docs/Tasks/Chat. Use this skill even when the request seems simple — gws can handle it in one command.

Generate, edit, or understand images using Google Gemini models. Use when asked to "generate an image", "create a picture", "edit an image", "modify this photo", "describe an image", "what's in this image", "analyze this photo", "image to text", "text to image", "explain this screenshot", "visual question answering", or any task involving image generation, image editing, or visual understanding.

Transfer context to a new focused session. Use when starting a new phase of work, handing off to a new session, or when context needs to be summarized for continuation. Triggers on "handoff", "transfer context", "start new session with context", or when the user wants to continue work in a fresh session.