skills

Delegate one or more coding tasks to background worker agent-CLI sessions (claude by default; codex, gemini, or any other agent CLI via profiles) and babysit them — spawn each in its own tmux pane or cmux workspace, send a prompt, poll/wait for completion, collect the result, then stop. Use when the user says "spin up a worker / sub-agent to do X", "run these in parallel", "fan this out", "delegate X while I keep working", "drive another claude / codex / gemini", or otherwise wants long-running tasks handled by separate agent instances you supervise. SAFETY: workers run in full-auto mode (e.g. claude --dangerously-skip-permissions — no per-action approval) in whatever directory you point them at — only use on repos/dirs the user trusts you to modify autonomously, and surface NEEDS_INPUT to a human rather than auto-answering.

Shape every reply to the smallest complete answer. Right-size to the ask — trivial asks get one bare sentence; bigger asks get a bottom-line verdict then the tightest fitting structure (table, steps, code, or a small ASCII diagram), with a leading index of headlines only when the answer runs ~4+ sections. Use on every turn to keep replies short and scannable.

Present a list of items one at a time for focused, interactive review — each with enough context to act on, then wait for the user before advancing. Use when the user wants to step through items one at a time, or hands over findings / tasks / suggestions / a file of items to review interactively rather than as a wall of text.

Maintain a persistent visual work tree in an Obsidian Canvas as you (and the agents you're driving) branch into sub-problems, decisions, and dead ends across sessions. Resolves which canvas to use per project, tracks a small active set (1–3 nodes you're working on now — the 👉 primary is the attach anchor), and keeps the layout tidy. Use when the user says "add to the tree/canvas", "branch off X", "focus on Y", "also working on Z", "mark X active", "done with X", "new topic X", "close / finish the X topic", "tidy the tree", "where was I", "what was I doing", "I lost my place", or otherwise wants to capture / navigate the shape of an ongoing investigation.

Publish HTML artifacts (mockups, demos, reports, prototypes) from a local folder tree to a Cloudflare-hosted site at ubuilt.dev, preserving the folder structure as URL paths, with per-folder access control — some folders public, others gated behind email one-time-PIN for named collaborators who need no Cloudflare account. Additive and multi-machine safe: every publish reconciles against live Cloudflare state (R2 objects + Access apps) and never clobbers what another machine pushed. Use when the user says "publish this artifact / mockup / prototype", "share this with <person>", "put this folder behind email auth", "publish to ubuilt.dev", "make this folder public / protected", "unpublish X", or otherwise wants to deploy an HTML tree to their Cloudflare artifacts site and control who can see each part.

Investigate a given topic against the codebase, docs, and (when relevant) the web, then — only if it surfaces a concrete action that can't be resolved right now — draft a lean, assertive GitHub issue and create it for the relevant repo via the gh CLI after the user confirms. Use when the user hands over a topic, idea, or question to look into and may want it captured as an actionable ticket — e.g. "look into X and open an issue if needed", "investigate Y", "should this become a ticket?".

Shape every assistant reply. Simple questions → terse direct answer, freeform. Complex / multi-topic / process answers → open with headlines + one-line summaries, optionally an ASCII visual (diagram, mockup, table, flow) when order or layout matters, then short detail sections that each say what was understood, how it was done, and (when non-obvious) why. Use on every turn to keep replies scannable and aligned.

Audit a plan vs an implementation. Build an interactive HTML checklist of end-user acceptance criteria, then spawn parallel agents to verify each item against the actual code. Use when validating whether a built feature matches its stated plan, stories, tasks, or subtasks — or when you need a copy/paste-friendly feedback artifact tied to the plan.

Save content from the current conversation to /tmp/<descriptive-kebab-name>.md, then return the absolute path and an outline (every header with a one-line summary). Use when the user says "save this to /tmp", "snapshot this", "dump the X part", "stash this", "save the discussion about Y", or otherwise asks to drop content from the chat into a quick scratch markdown file.