Typing 'ultrawork' or 'ulw' flips the agent into a strict failing-first, evidence-driven engineering mode with mandatory cleanup receipts.
Investigates LazyCodex/omo-codex/Codex CLI defects with runtime and source evidence, then files a high-signal GitHub issue or PR in the repo that owns the bug.
Explains any senpi `Tip:` line in depth by querying the live tip list, checking your own settings, and reading the real feature source code.
Triggered by the keyword 'ultrawork' or 'ulw', it forces a failing-proof-first, real-surface-QA, cleanup-receipt workflow before anything can be called done.
Teaches Claude the correct, up-to-date patterns for creating, reading, writing, concatenating and subsetting AnnData (.h5ad/zarr) objects in single-cell workflows.
Runs a named team of cooperating Codex workers under the main session as leader, with durable on-disk team state.
Turns your full claude-mem project timeline into a narrative "Journey Into [Project]" development-history report.
Debugs a LazyCodex or upstream Codex CLI defect in a throwaway workspace and ships it as a fork PR or a patch-embedded verified-fix issue.
Strips AI-generated code slop from branch changes category by category — but only after regression tests lock the current behavior.
Turns a project's full claude-mem timeline into a serial, chapter-per-ISO-week narrative built by a chained sequence of subagents.
Splits a task into the smallest mergeable PRs, builds each in its own git worktree with evidence-backed manual QA, then loops on CI and review-bot gates until merged and cleaned up.
Grounds itself in your codebase, asks only the questions exploration can't answer, and after your explicit okay writes one decision-complete work plan a worker can execute with zero follow-up interview.
A 6-phase refactoring protocol: clarify intent, build a codemap, gauge test coverage, plan atomic steps, and verify after every change.
Teaches Claude how to design, submit and retrieve protein experiments through the Adaptyv Bio Foundry API and Python SDK.
Look up precomputed AlphaGenome Atlas AVI scores and call the AlphaGenome model on demand to rank and mechanistically interpret non-coding regulatory SNVs.
Searches scientific literature through the BGPT MCP server and returns 25+ structured fields extracted from full-text papers.
A careful runbook for moving legacy NanoClaw and Claude-native memory into the shared memory/ tree with quarantine, reporting, and rollback.
Tells Claude to call real external APIs directly while the OneCLI HTTPS proxy injects stored credentials at the boundary.
Scans your API routes and auto-generates test suites covering auth, validation, error codes, pagination and rate limits.