| name | ultrawork |
| description | Orchestrate large Claude Code tasks with planning, todo discipline, subagent delegation, parallel execution, and targeted verification. Does not stop until the task is done or explicitly blocked. Activate when #ulw appears anywhere in the user message. |
| argument-hint | [goal] |
| disable-model-invocation | false |
| allowed-tools | Read, Edit, MultiEdit, Write, Glob, Grep, Bash, Task |
Execute this workflow for: $ARGUMENTS
Use this when the task is larger than a quick edit or when the user wants rigorous execution.
Phase 0 — Intent gate:
- If the user only asked a question, answer directly and do not force the full workflow.
- If the task is trivial (single file, obvious fix), execute directly without ceremony.
- For everything else, proceed to phase 1.
Phase 1 — Context and planning (invest tokens here to save rework later):
- Restate the goal in one sentence.
- Read repo-local
CLAUDE.md if it exists, .claude/state/ if it exists, and the minimal repo files needed to discover canonical commands.
- If
.claude/state/memory/index.md exists, read it to leverage project knowledge from past sessions.
- If
.claude/state/briefings/ has recent entries, read them for continuity.
- If
.claude/state/current-task.txt is empty or stale, create or refresh a task note with /omo:kickoff-task.
- Initialize Boulder for cross-session persistence:
bash scripts/boulder-init.sh "<goal>"
- Create a todo list before editing. Keep it updated as work progresses.
- Parallel discovery: Run
deepsearch + repo-librarian (or Explore) simultaneously for broad context. Use planner-sisyphus when the task is still fuzzy after initial discovery.
- If the plan has 3+ steps, run
critic to verify the plan is executable before starting. Do not skip this — catching plan flaws now is cheaper than fixing them mid-execution.
Phase 2 — Execution (parallelize aggressively):
- If the task is a bug, bring in
bug-hunter + deepsearch in parallel before editing. If verification is expensive or unclear, bring in test-commander.
- If the work splits cleanly into independent slices, use
/omo:spawn or delegate up to 5 slices in parallel via Task with run_in_background=true. If the repo is a git repo and the change is much larger, suggest built-in /batch instead.
- Slice size limit: Each delegated slice should target ≤3 files or ≤3 edit sites per file. Never paste entire file contents into a Task prompt — let the agent read files itself.
- Docs vs code: Documentation files (
.md, guides, READMEs) should be delegated to docs-keeper or edited directly, not to build-integrator. Reserve build-integrator for code files only.
- Prefer parallel over sequential: When slices are independent, always dispatch simultaneously. Two agents in parallel > one agent doing both sequentially.
- Implement the smallest coherent slice first. Re-plan after the first slice if reality changed.
- After each meaningful edit cluster, run targeted verification immediately.
Phase 3 — Verification and completion (multi-perspective, never single-angle):
- Run
/omo:ship-check AND /omo:diff-review together — dispatch both in parallel for comprehensive final verification.
- If verification fails, fix and re-verify. Do not stop at partial completion. Escalate to
oracle after the first failed fix attempt.
- If the task pauses before completion (context limit approaching), run
/omo:handoff.
- On successful completion, finalize Boulder:
bash scripts/boulder-complete.sh
- Before finishing, provide:
- outcome
- files changed
- verification results
- remaining risks
Specialist delegation (prefer specialists over inline analysis — they are deeper and can run in parallel):
oracle for architecture decisions or after the first failed fix attempt.
bug-hunter + deepsearch in parallel for debugging and failure triage.
test-commander for verification strategy.
build-integrator for code implementation slices (not docs).
deepsearch for comprehensive codebase search (run alongside other specialists).
critic for plan review before execution (mandatory for 3+ step plans).
security-auditor for any changes touching auth, input handling, or external APIs.
vision for screenshot or image analysis.
docs-keeper for documentation and guide file changes (including new content, not just cleanup).
Guardrails:
- Do not spend turns building abstractions for the workflow itself.
- Prefer repo-local skills and agents over user-global ones.
- If an MCP server is missing, continue with native tools rather than blocking.
- Do not stop until all todos are complete and verified, or a hard blocker is found.
- Do not expand scope beyond the original goal without explicit user approval.