Think a feature through with /dev:design, then plan, build, verify and land code changes through one pipeline — /dev:fix, /dev:task, /dev:feature, /dev:build and /dev:deliver — with approval gates, sequential implementers, a subtractive prune pass and stacked PRs. On request, /dev:review fixes what nobody would argue with and posts the rest that survived a refutation pass as inline PR findings, /dev:fix-pr-comments acts on the unresolved PR threads and asks in chat where a fix needs a choice, /dev:review-and-fix loops the two until review calls the branch finished, /dev:explain-pr explains a PR or a whole stacked chain in plain language, and /dev:update-packages updates every dependency in verified batches and flags the deprecated or abandoned ones. Project-specific checks and git flow come from the project's own verify and git-workflow skills, which /dev:setup and /git:setup write
14 skills
Git operations agents get wrong in ways that silently lose work — resolving merge, rebase and cherry-pick conflicts by both sides' intent, a hook that blocks unnamed stashes, pops by no ref or by position, plain force-pushes and, where a project opts in, pushes to its protected branches, and /git:setup, which writes the project's git-workflow skill and .claude/git.json
2 skills · hooks
One-way Telegram notifications: let the agent reach your phone when you are away from the terminal
2 skills · 1 MCP server
Multi-account Google Drive for agents: browse, read Docs as Markdown, download, upload, edit and trash files — each permission switched per project, several Google Cloud apps at once, never deletes permanently
1 skill · 1 MCP server
Multi-account Gmail for agents: search, read as Markdown, drafts with attachments, label rules — configured per project, several Google Cloud apps at once, never sends
1 skill · 1 MCP server
Protocol for effective web research: engine selection, query hygiene, graded negative verdicts and primary-source verification
1 skill · 1 MCP server
Mechanical safety guards for agents: hooks that refuse a dangerous command and tell the agent what to do instead. Guards: no inline node/bun eval (-e, -p), which turns any library that forks with execArgv into a fork bomb; no package install inside a git submodule whose packages are workspaces of the superproject, which shadows its dependencies; no command that prints a secret value into the transcript (an env dump under secrets run, a keychain read, the store token, bws secret get)
hooks
Hand the work to a fresh agent: a summary of the session that points at existing artifacts, then a new Paseo agent (or a background claude) started with it
1 skill
Find a name for a product, app or site over several rounds: candidates in chosen directions, domain checks, conflict checks, and a record in the repo so no feedback has to be repeated
1 skill
Playwright MCP browser automation, run on bun instead of npx — the one browser every skill in this marketplace depends on
1 MCP server
Retrospective on how agents worked in a project: root cause from Claude Code and Codex transcripts and PR reviews, verdicts on the previous retro, and proposed changes to the environment
1 skill
Collect API keys and tokens through a masked dialog the agent cannot read, and hand them to scripts without exposing them — no secret ever pasted into the chat
1 skill
Find the best product to buy against your criteria: an accepted spec first, then a wide longlist, parallel deep dives and a ranking with the best obtainable price and a link to buy every pick
1 skill
Revise an existing skill so the agent behaves as intended, without the skill swelling — evidence from real transcripts first, knowledge grows freely, procedure asks
1 skill
/wait-what: the last message did not land, so the agent pitches it again plainly, with context
1 skill
Read WhatsApp chats and attachments straight from the macOS desktop app's local database — list chats, read or search messages, find and copy files. Read-only, no API, no linked device
1 skill