Ultrawork

What it does

When a prompt contains ultrawork or ulw, this directive is injected into the conversation and reshapes how the agent works — evidence over narration:

• One-time tier triage (LIGHT/HEAVY): new modules, auth/security, external integrations, DB migrations, concurrency, cross-domain refactors force HEAVY; you may ratchet up mid-task but never down.
• Binding goal registration: create_goal records the deliverable, tier justification, success criteria with literal commands and binary PASS/FAIL observables, and an explicit "when to stop" line.
• Durable notepad: an append-only mktemp markdown file holding the plan, file:line findings, RED/GREEN captures, and evidence paths so state survives context compaction.
• PIN → RED → GREEN → SURFACE → CLEAN loop: capture the failing proof before any production code, make the smallest change, re-prove on the real surface (curl, Chrome/agent-browser, xterm.js web terminal, OS-level GUI automation), then tear down every spawned server, tmux session, container, or temp dir and log a cleanup receipt.
• Parallel + delegation discipline: batch independent lookups into one eval cell; fan out task subagents or a team_create team only when write scopes are disjoint.
• Triggered verification gate: reviewers (momus/metis) are summoned only when a plan file exists and the tier is HEAVY or the user demanded rigorous review; otherwise a recorded self-review.

Who it's for

• Developers tired of "should work" claims with no captured proof
• Teams handing high-blast-radius changes (auth, payments, schema migrations) to an agent
• Advanced oh-my-openagent / omo-senpi users standardizing evidence-based QA and multi-agent delegation

Examples

1. ulw add email format validation to the login API → write the failing test (RED), implement minimally, capture curl -i showing 400 status + body, log evidence paths and an atomic commit.
2. ultrawork build the payment webhook integration → auto-classified HEAVY: 3+ criteria (happy, boundary, adjacent regression), per-criterion scenario and evidence, plus reviewer approval if a plan file exists.
3. ulw fix the broken colors and CJK width in the TUI → prove it with a screenshot of a real xterm.js-rendered terminal instead of tmux capture-pane, then record the pty/tmux teardown receipt.

Install the skill I found on Claude Skill Mart.
Copy the packages/omo-senpi/skills/ultrawork folder from the GitHub repo code-yeongyu/oh-my-openagent into my ~/.claude/skills/code-yeongyu-ultrawork/.
When it's done, tell me in one line what this skill can do.