Writing & Editing

Put work on the correct branch before the first edit. Finds the base branch, starts a well-named branch from the current remote tip, and sets up one worktree per agent for parallel work. Use at the start of any change to a tracked file, such as 'start on issue 42', 'fix this typo', 'set up branches for these three tickets', or before you start sub-agents that edit code. Not for opening or merging a PR (github-pr-merge) or for history rewrites and other risky git commands (git-safety).

Approval gate for git commands that destroy work or skip checks: force-push, --no-verify, signing bypass, reset --hard, clean -f, discarding uncommitted edits, branch or tag deletion, rewriting pushed commits and push --mirror. Each one needs the user's explicit approval at that moment; an earlier approval never carries over. Also covers a secrets check before a commit and what to do when a secret is already committed. Use before any of these commands, or when asked to 'force push', 'skip the hooks', 'throw away my changes' or 'delete that branch'. Not for normal branching (git-branching) or the PR flow (github-pr-merge).

Create or top up a project's documentation set: list what exists, let the user choose which document types to add, collect one shared fact sheet from the code, write the documents (in parallel when there are several), then check them against each other and against the code. Use when asked to 'bootstrap the docs', 'start the documentation', 'we have no README or ARCHITECTURE', or 'fill in the missing docs'. Not for the per-commit update of existing documents (documentation-sync), and not for gener8v pipeline documents (gener8v:brownfield and gener8v:planning make those).

Write or improve Realms of Code realm guides with practical triggers, concrete actions, evidence and fantasy flavor. Use when authoring the charter itself, not when applying its roles to ordinary software work.

>- Use when planning or implementing personalization or A/B testing with Croct: creating an experience (audience, slots, content, priority, scheduling, drafts), adding an experiment (A/B/n, primary goal ID, traffic allocation, variants, cross-device consistency), wiring the code it needs (slots, `croct.track('goalCompleted', {goalId})`, `orderPlaced`, `leadGenerated`, `userSignedUp`), server-side feature-flag experiments, redirect experiments, personalization recipes (UTM campaigns, ABM, search terms, location, free shipping, interests, blog readers, leads, customers, organic intent, HubSpot forms), or reading the experience, experiment, application, or workspace dashboards (probability to be best, potential loss, "Experiment in progress", "Recommended winner", "No clear winner", "Track your goal"). Phrasings: "personalize the hero for LinkedIn visitors", "A/B test this CTA", "which variant won?", "why does my experiment show no conversions?".

>- Use when writing, reading, reviewing, or debugging Croct's Contextual Query Language (CQL): audience criteria, `croct.evaluate()` / `useEvaluation` / `<Personalization query>` / Next.js `cql` tag queries, Evaluation API `query` bodies (`/client/web/evaluate`, `/external/web/evaluate`), and dynamic content values (`"type": "dynamic"`, `"expression"`). Covers context variables (user, session, page, location, campaign, cart, device, browser, context, now/today), data types and conversions, accessors and selectors (`'s`, `property X of`, `[1st]`, `[*]`, `[...]`), tests (`is`, `matches`, `is like`, `contains either`, `is between`, `is defined`, `is returning`), quantifiers (`some ... satisfies`), macros, modifiers, event expressions (`user has viewed a page`, `user has abandoned a cart`, `user has completed goal`), audience estimation, and recipes. Triggers: "write an audience", "target mobile users in New York", "CQL syntax error", "audience never matches", "unavailable estimate".

>- DỌN RÁC Ổ ĐĨA (Windows): quét thư mục home ở chế độ CHỈ ĐỌC (chỉ đo dung lượng, không mở nội dung file), tìm dữ liệu tồn đọng từ quá trình code (cache npm/pip/uv/pnpm, Temp, cache trình duyệt/VS Code, node_modules, .next, venv, lịch sử phiên Claude Code/Codex, Docker/WSL, model AI), rồi dựng ARTIFACT "Bản đồ ổ đĩa": treemap dung lượng + danh sách đề xuất chia 2 nhóm "Có thể xóa an toàn" / "Cần xem lại trước khi xóa", tick chọn rồi Copy script PowerShell dùng lệnh clean chính thức của từng công cụ. Ghi ~/disk-report/sizes.csv và lần sau so sánh, đánh dấu mục tăng dung lượng. KHÔNG tự xóa gì. Dùng khi người dùng nói "dọn rác", "dọn ổ đĩa", "ổ C đầy", "hết dung lượng", "máy đầy", "quét dung lượng", "bản đồ dung lượng", "disk cleanup", "giải phóng ổ cứng".

Generate a polished, print-quality Typst technical document (with diagrams, tables, quoted evidence, and hyperlinked references) when the user asks for a detailed writeup, investigation report, or architecture doc — a strong alternative to Word/.docx.

Audit and fine-tune an existing agent skill so it triggers when it should and hands off cleanly to its siblings. Reads the real SKILL.md and references, checks the description, overlaps with sibling skills, stale names and counts, handoff contracts and command routing, makes minimal evidenced edits, validates them with a script, and writes near-miss test cases. Use whenever the user wants to fine tune, improve, tighten, debug or review a skill, says a skill is not triggering or fires on the wrong requests, wants two skills to work together or split cleanly, or pastes a SKILL.md or repo link and asks what is wrong with it, even if they never say "audit". Not for creating a skill from scratch (use draft-skill) or from a list of prompts (use prompts-to-skill).

Write and publish a post to a personal Jekyll + GitHub Pages blog — use when the user says \"blog this\", \"write a post\", \"push to the blog\", or \"/blog\

Generate coordinated editorial watercolour blog art.

Generate coordinated editorial watercolour blog art.

Turns a vague or half-defined idea into a spec (problem, users, user stories, requirements, edge cases and acceptance criteria) through a round-based interview. Use it when asked to write a spec, define a feature, clarify requirements or go from an idea to something implementable.

Use when documenting how a music hardware device is controlled over MIDI (CC, NRPN, SysEx) or USB so a controller, editor or librarian can be built: finding official docs, capturing traffic from the user's own device, decoding patch dumps, and writing a verified MIDI implementation document.

Complete coding, building, research, and writing tasks with focused autonomy, clear scope, honest verification, and efficient recovery. Use for substantial work or when asked to follow our routine.

Use when reviewing, writing or sizing a Thanos deployment (Helm values, manifests, component flags, alert rules, cache config) or diagnosing Thanos read path saturation - store gateway gate pinned, slow global queries, first query after idle endless, Receive OOM, compactor alerts that never or always fire, cache sizing for Redis, Memcached or Dragonfly.

>- Run the Mampu AI work loop on any work-shaped request — a bug to fix, a feature to build, a doc to write, a client ask to handle, an investigation, a PR to ship, a deploy. Finds or creates the team task, reads the matching internal records and help pages, opens the reply with the four-line status block, and closes the task out with a debrief. Not for a plain question or a read-only look.

>- Generates a UNIT TEST suite for embedded firmware .c/.h files, following Maurício Aniche's specification-based testing (partitions → boundaries → cases → coverage). Detects the environment: Zephyr uses twister + ztest + FFF on native_sim; ESP-IDF uses native Unity in a host build + gcovr; anything else follows the project's existing test framework (Unity, CMock, Ceedling, GoogleTest…) or stops and asks if there is none. ALWAYS use it when the user asks to generate/write/create unit tests for .c/.h modules — e.g. "write unit tests for motor_ctrl.c", "create unit tests for the uart_driver", "cover these .c/.h files with tests", "test this function with partitions/boundaries", "95% coverage on this module", or in Portuguese "gera os testes unitários de motor_ctrl.c", "cria unit test pro uart_driver", "cobre esses .c/.h com testes". Triggers without the word "skill" and without naming the framework. Do NOT use for integration, e2e, HIL, system or Robot Framework tests; nor to scan the project for targets (o...

Turns a repo's code review history into numbered coding standards, each rule with its why and the review comments behind it, then reviews diffs against them and keeps them lean. Use when writing or auditing coding standards or conventions from review history (mine, write), reviewing a diff, branch or pull request against standards (review), pruning standards that agents already follow unprompted (prune), or checking standards for drift, such as rules whose files are gone or prune data measured on an older default model (doctor).

把播客单集链接转成按章节整理的完整文字稿（本机 CapsWriter-Offline 语音识别，产出 chapters.md，非摘要）。用户给出播客/音频链接要求转录、转文字稿、整理成文，或问"那个播客转文字的 CLI 叫什么"时使用。本地音频文件（无链接）转写走 FunASR-Nano，见文内参考。

Pull what is due and what is coming from Canvas through the browser. Use when Srijan says "what's due", "canvas", "this week", "what do I have", "any tests coming", or at the start of a planning session. Reads the Canvas dashboard and each course's upcoming items through Claude in Chrome, lists them by date, adds test dates to Google Calendar, and writes the week into the notes.

Write or revise school writing to the actual rubric. Use when Srijan has an essay, short answer, DBQ, lab write up, reflection, or any graded writing. Finds the prompt and rubric first (Canvas, handout, teacher slides), builds an outline that hits every rubric line, drafts in his voice, then grades the draft against the rubric before he sees it.

Set up and manage local development databases on the user's Mac through TableStudio. Use when a project needs a local Postgres, MySQL, ClickHouse or SQLite database, or a DATABASE_URL for one; when a local database refuses connections or is stopped (ECONNREFUSED, "connection refused", "could not connect to server", "Can't connect to MySQL server"); when the user wants a scratch or test database or a safe copy to experiment on; or when a PostgreSQL extension (pgvector, PostGIS, pg_trgm) needs installing.

Find and fix slow database queries with TableStudio, proving each fix with measurements, and check a database's overall health. Use when a query, page or endpoint is slow, when asked which queries are slow or what's wrong with the database, when adding, reviewing or dropping indexes, when an ORM query needs tuning, or when deciding whether a SQL rewrite is faster.