Initialise, adopt, maintain, audit, or hand over a vendor-neutral workspace: the repository's files, not any provider's chat history or hidden memory, are the durable record of an AI agent's work, so the same subject resumes from another machine or another harness. Memory Bank context files, MADR decision records, a Markdown or Beads task backend declared in workspace.yaml, session records and approval packets. Ships the standard, a template, scripts to create, validate and migrate a workspace, and Claude Code hooks that enforce the session record.
1 skill · 1 command · hooks
Rewrite drafts, articles, emails, Slack messages, proposals, or LLM-generated copy into Joel Azemar's voice so it reads human, not AI-generated: picks the register (public company copy or Joel's own messages), enforces the no-dash and no-invented-facts rules, cuts hedges and empty intensifiers, and returns paste-ready text plus a short note of what changed.
1 skill
Drive the end-to-end execution of any coherent piece of code work in a GitHub-hosted project: size-scaled design pass, issue, Kanban, branch, implement, internal code and security review loops, validation gates, atomic commits, live verification, PR, automated-reviewer loop, review-thread triage, merge per project policy, optional release, and persisting what was learned. Project-specific values come from the project's own workflow doc.
1 skill
Rewrite a draft prompt before sending it so the answering model tests the asker's frames instead of inheriting them: inventories every injected frame, keeps real constraints and evidence, quarantines the rest as hypotheses, states the question neutrally, and adds instructions that force competing framings and odds-backed conclusions. Returns the rewritten prompt plus a table of what changed.
1 skill
Counts down until the chat's prompt cache goes cold, above the prompt. When a big chat goes cold, offers a handoff to a fresh chat. Keep warm pings the cache in the background so it does not go cold
Reads each prompt, picks the effort level it needs with Ben's four-question framework, and offers Switch to that level, or sets it by itself in auto mode
At 250k tokens, offers Refresh: writes a handoff of the whole chat, shows the steps with progress bars, then copies it or starts a new chat with it. Logs every run to ~/.claude/handoffs/handoff.log and saves a draft first
Helper Desk: a purple task progress bar above the prompt that moves as each to-do item is done, and a side pane with one live card per subagent
One interface for terminal multiplexers ($.mux): floating panes, panes beside the chat, a floating editor, and the active theme's colours. Zellij is supported; with none, every call answers null so each mod falls back on its own. Other multiplexers (tmux) plug in as a backend
One row above the prompt: how much of your 7-day and 5-hour usage windows is left, and when each resets. LOW under 15%. Reads the engine's own figures, so it costs no tokens; draws nothing off a subscription