Productivity

Use when executing an approved plan end to end without a human watching. Typical cues are "here's the plan, go", "work on this overnight", "run autonomously", "I'll be back in a few hours", or a long session started before a meeting. Covers what the agent may decide alone, the four checkpoints it must never pass, keeping work durable (commit and push often), delegating to subagents, and how to leave blockers so the work is reviewable when the human returns.

Use when a PR stage or feature is finished and before its PR opens; when files, packages, routes, commands, env vars, APIs or conventions were added, renamed, moved or removed; or when asked to "update docs", "sync docs", "refresh the README/AGENTS.md", or told "docs are stale". Audits every doc against the code, fixes drift, regenerates generated docs, and commits the result to the current branch.

Arm, run and disarm an unattended away run (a night or a few hours). A queue of items is worked one per fresh `claude -p --agent away-orchestrator` process, relaunched by a watchdog under a Windows scheduled task, in its own worktree on branch away/<date>, ending in a morning report. Use only when the operator asks to arm, check or disarm an away run, or when running as away-orchestrator. Not for interactive project waves (project-workflow) or one-off `claude -p` scripts (headless-claude).

Mine recurring user corrections and workflow friction from agent session transcripts, then convert mechanically detectable patterns into tested rules — use for \"mine rules\", \"what rules should we create\", \"turn this friction into a rule\", repeated corrections, or \"/ffs\".

Splits an approved technical plan into small tasks, ordered by dependencies, with affected files, per-task verification and parallel groups. Use it when there is a plan and someone asks to break it down into tasks or get it ready for implementation.

Creates or updates a project's context files (AGENTS.md, docs/ARCHITECTURE.md with its non-negotiable principles, CONTRIBUTING.md, ADRs, SECURITY.md and PR template) through a guided interview. Use it when starting a project, when adopting agents in an existing project, or when asked for a "constitution", project principles, architecture rules or documentation for agents.

ALWAYS load at the start of every session and before any task with work in it. Captures every task and sub-task into a backlog kept on the user's own computer, with area, ETA and session, and updates it every turn so no work is forgotten in long or parallel sessions.

Coordinate an engineering task across authorized agents with explicit ownership, bounded handoffs and evidence. Use when independent implementation, review or verification adds value; handle ordinary single-agent work directly.

Safe Git and GitHub routine for this repo (TTT-b/stremio-ai-translator). Use before pulling, committing, pushing, tagging, or opening a pull request.

Choose the cheapest model that does the job and keep token use low. Use at the start of any task, before planning, delegating to a subagent, or reading many files.

Analyze a codebase and recommend Claude Code automations (hooks, subagents, skills, plugins, MCP servers). Use when user asks for automation recommendations, wants to optimize their Claude Code setup, mentions improving Claude Code workflows, asks how to first set up Claude Code for a project, or wants to know what Claude Code features they should use.

>- Connect a Mampu AI teammate to the company Lark and docs and onboard them into the work loop. Use when someone says connect me, onboard me, sign me in, set me up, join the work loop, "I want to be connected to the company Lark / docs", "how do I get started", or when the session-start context says they are not connected or their sign-in lapsed. Guided and conversational: you run the steps, they only approve a browser sign-in.

State manager for multi-dimensional computational experiments. Iterate a parameter grid over many rounds, snapshot declared critical files each round, audit every round (params, hashes, timestamps, status), and roll back to any round non-destructively. Use it when a workflow runs hundreds of rounds across multiple parameter dimensions and needs per-round snapshots, full audit trail, and arbitrary rollback - e.g. longitudinal studies, parameter scans, or symbolic/ numerical sweeps calling external solvers (Mathematica, GAP, SymPy, custom scripts).

Investigate a bug, regression, flaky behavior, or unclear backend failure with an evidence-first debugging workflow before jumping into fixes. Use when Codex should establish a reproducible feedback loop, rank hypotheses, inspect the real failing path, and only then propose or apply changes.

>- Ships an app clone live on the user's own domain: a preflight gate (tests green, parity must-haves done, rebrand sweep clean, listing linted, legal pages up), production database and env vars, the host, DNS records for the domain and for email, Stripe live mode, OAuth redirects, monitoring, and mobile builds to TestFlight and Play. Use when the user says "deploy it", "ship it", "put it live", "connect my domain", "go to production", "publish the app", or after /replica-launch.

>- Builds the backend of an app clone: auth, database migrations and access rules, payments with Stripe, email, background jobs and third-party integrations through official APIs only, plus a security checklist. Use when the user says "add login", "set up auth", "wire up the database", "add payments", "connect Stripe", "add Google Calendar", "send emails", "backend for my clone", or when /replica-build is running on fake data.

Use when a change needs an independent code review before it moves on - a finished task, a feature branch, a pull request, or work-in-progress changes since a given ref - in any programming language. Pins the exact diff, gathers intent and constraints, and dispatches a read-only reviewer that returns separate verdicts and evidence-backed findings.

Turn a pasted task or ticket description into a concrete implementation plan for this repo.

Work with Projects in the Obsidian PARA vault and their pairs in the task app — list the active projects, create a project (folder plus a project in the task app), summarize where a project stands, or add a task to a project in the task app (Todoist, for example). Use when the user says "list projects", "show projects", "what am I working on", "create a project", "new project", "project overview", "project status", "add a task", "какие проекты", "список проектов", "покажи проекты", "что у меня в работе", "заведи проект", "новый проект", "саммари по проекту", "обзор проекта", "что по проекту", "статус проекта", "где я по проекту", "добавь задачу", "заведи задачу".

Work with the inbox of the PARA vault, which lives in the task app's inbox — capture a raw thought into it, list what is sitting there unprocessed, or process captures one by one into their proper PARA folder. Use when the user wants to capture a note, mentions their task app's inbox (Todoist «Входящие», for example), or says "capture this", "add to the inbox", "what is in the inbox", "process the inbox", "запиши", "закинь в инбокс", "что в инбоксе", "покажи инбокс", "разбери инбокс", "разгреби инбокс".

Work with Archive in the Obsidian PARA vault — put a finished or abandoned project (folder and its project in the task app, such as Todoist) or a note into the archive, list what is already there, or restore something back into an active folder. Use when the user says "archive this", "move to archive", "close the project", "what is in the archive", "show the archive", "restore from the archive", "заархивируй", "перенеси в архив", "закрой проект", "что в архиве", "покажи архив", "верни из архива", "восстанови проект".

Save something to the right place in the user's AI file system. Use when they say "remember this", "save that", "note that I prefer...", "from now on...", run /ai-file-system:remember, or at the end of a task when something lasting was learned.

Build the user's AI file system, one organized folder of plain-text files (AGENTS.md, CLAUDE.md, CURRENT.md, context files and workspaces) that Claude and Codex read at the start of every task. Use when someone says "set up my file system", "build my OS", "set up my AI folder", "organize my AI", or runs /ai-file-system:setup. Interviews them, optionally imports their ChatGPT or Claude history, builds and verifies the files, and shows them how to use it every day.

Read-only root-cause investigation for one explicitly assigned GitHub Issue; returns a durable diagnosis to the Issue and Scheduler, without editing code, opening a PR, reviewing a PR, or creating tasks.