Use this skill when adding, registering, or theming a new ISpect log type / log key (e.g. "add a ws-state log", "register a new log key", "new trace category", "my custom log has no color/icon"). It is the checklist for wiring a key end to end so nothing is missed — core registration, console color, UI color, icon, title, and localized description.
Create or edit AgentSync configuration — AGENTS.md, rules, skills, commands, subagents, settings, MCP servers, hooks, or per-tool configs. Use this skill when adding a rule, creating or scaffolding a skill, writing a slash command, defining a subagent persona, editing permissions, configuring an MCP server, setting up the `.ai/src/` directory, or running `agentsync sync` / `add` / `customize` / `resolve` / `simplify` — even when the user does not name "AgentSync" explicitly but is editing files in `.ai/src/`, `.claude/`, `.cursor/`, or another tool-config directory.
Rewrite text so it reads like a real human wrote it, or generate long-form prose (blog posts, essays, articles, newsletters, social posts, op-eds, short stories) in that same plain, grounded voice. Strips common AI tells, including em dashes, semicolons, framing colons, inflated vocabulary (utilize, leverage, delve, tapestry, robust), contrastive negation, significance inflation, forced triads, sycophantic openers, and chatbot closers. Trigger whenever the user asks to humanize, de-AI, or de-slop text, even with casual phrasing like "make this less AI-sounding," "fix the AI vibe," or equivalent phrasing in any language. Also trigger when the user asks Claude to write a prose deliverable they will publish or share, like an article, blog post, essay, newsletter, or short story. Do not use for code, casual chat replies, or technical documentation.
Scan recently written or edited code and remove comments that do not earn their place — narration, step markers, task/PR references, commented-out blocks, decorative banners. Use this skill after finishing a non-trivial code change, when reviewing a diff for comment hygiene, when the user mentions noisy or redundant comments, or proactively before declaring a coding task complete. Defer to the `comments` rule for what makes a comment worth keeping.
Create a well-structured git commit for staged or unstaged changes — analyze the diff, match the project's existing commit style, write a clear message, and create the commit. Use this skill when the user asks to commit, save changes, check in code, write a commit message, prepare changes for push, or wrap up a piece of work — even when they don't say "commit" explicitly (e.g. "let's save this", "ship it", "сохрани изменения").
Investigate and fix bugs, errors, or unexpected behavior systematically — reproduce, locate, understand the root cause, fix it, and add a regression test. Use this skill when the user reports a failure (test, runtime, build, CI), shares a stack trace or error message, says something doesn't work, asks why something is broken, or asks for a fix that requires diagnosis — even when the word "debug" is not used (e.g. "this is broken", "почему падает", "не работает X").
Design, debug, and refine prompts for LLM-based coding tools, agents, and pipelines. Use this skill when writing or rewriting a system prompt, agent persona, slash-command body, rule, or skill description; tuning a prompt that produces vague, incomplete, off-format, hallucinated, or over-/under-triggering output; choosing among techniques like XML tags, few-shot examples, role-setting, reasoning steps, or recap; debugging a model that ignores instructions, leaks information, or over-engineers solutions; or designing the persona, autonomy posture, or tool-use rules of an autonomous agent — even when the user does not explicitly mention "prompt engineering" or "system prompt". Not for AgentSync file structure, frontmatter, or scaffolding new rule/skill/command files — use the `agentsync` skill for that.
Restructure existing code without changing its behavior — reduce duplication, improve naming, simplify complex logic, extract helpers, split overgrown functions, untangle dependencies. Use this skill when the user asks to refactor, clean up, simplify, DRY out, rename, extract a helper, split a file, or make code "nicer" — including when they describe code-quality concerns without using the word "refactor" (e.g. "this is messy", "приведи в порядок", "это можно сделать чище").