Create a plan from a brief:
omc ultragoal create-goals --brief-file plan.mdOr with explicit stories:
omc ultragoal create-goals --brief "ship the migration" \ --goal "Schema::Add new columns" \ --goal "Backfill::Backfill rows in batches" \ --goal "Cutover::Drop old columns and switch reads"The default mode is
aggregate(one Claude/goalcovers the run). Pass--claude-goal-mode per-storyif you want each story to have its own/goal.Multi-repo workspaces / parallel sessions: when several Claude sessions in the same workspace need to run
/ultragoalconcurrently, pass either--plan-id <stable-id>or--auto-plan-idso the plan is written to.omc/ultragoal/plans/{planId}/instead of the shared single-plan path. Without that flag, two sessions creating goals would clobber each other.--auto-plan-idderives{epochMs}-{slug}from the brief title. Then thread the same--plan-id <id>through every subsequent subcommand in that session. Useomc ultragoal list-plansto enumerate available planIds when needed.Start (or resume) the next story:
omc ultragoal complete-goalsThis prints a model-facing handoff. The active Claude agent must read it and:
- Confirm/Set the active
/goalcondition in this session. - Work the story.
- When the story is complete (and for the final story, after the full quality gate), share back a snapshot of the active
/goalstate and callcheckpoint.
- Confirm/Set the active
Checkpoint a story:
omc ultragoal checkpoint --goal-id G001-... --status complete \ --evidence "tests/files/PR evidence" \ --claude-goal-json '{"goal":{"objective":"...","status":"active"}}'For the final story, also pass
--quality-gate-jsoncontainingaiSlopCleaner,verification, andcodeReviewevidence (all clean).If the final review is not clean, do NOT mark complete. Record blockers:
omc ultragoal record-review-blockers --goal-id G00X-... \ --title "Resolve final code-review blockers" \ --objective "Fix the listed review findings and rerun final gates" \ --evidence "<the review findings>" \ --claude-goal-json '{"goal":{"objective":"...","status":"active"}}'This appends a new blocker story and keeps the Claude
/goalactive.Inspect state at any time:
omc ultragoal status
Ultragoal
Durable multi-goal workflow that persists plan/ledger artifacts under .omc/ultragoal and prints Claude /goal handoff text for the active session
Ultragoal by salacoste · 690b151
npx skillmds@latest add salacoste/ultragoal File contents
---name: ultragoaldescription: Durable multi-goal workflow that persists plan/ledger artifacts under .omc/ultragoal and prints Claude /goal handoff text for the active session---<Purpose>Ultragoal breaks a brief into an ordered set of goals, records start/checkpoint/blocker/failure events in a durable append-only ledger, and tells the active Claude agent how to drive the Claude Code `/goal` slash command alongside the plan. It does not — and cannot — mutate Claude `/goal` state from the shell; it persists durable repo state and prints a model-facing handoff that the active agent must act on in-session.</Purpose><Use_When>- The user wants a durable, repo-native way to track an ultragoal across multiple Claude sessions or worktrees- The work is large enough to warrant multiple ordered "stories" with attempt counts and per-story evidence- The user wants the final completion gated behind ai-slop-cleaner + verification + $code-review- The user wants the active Claude `/goal` directive coordinated with the ledger so that a session restart does not lose progress</Use_When><Do_Not_Use_When>- The task is a single small change — use direct delegation or `ralph` instead- The user wants the assistant to literally invoke `/goal` itself from the shell — that is not possible; `omc ultragoal` only writes artifacts and prints handoff text- The user wants a planning-only artifact with no execution loop — use `plan` instead</Do_Not_Use_When><Why_This_Exists>Claude Code `/goal` is a session-scoped Stop hook: it blocks the session from stopping until a condition holds, and auto-clears on success. That is a great single-session execution primitive, but it loses state across sessions and does not by itself enforce a final review gate. `omc ultragoal` adds a durable plan, ledger, and gating layer so a long multi-step initiative can survive session restarts, fresh worktrees, and review iterations while still leveraging Claude `/goal` to keep the active agent focused.</Why_This_Exists><How_To_Use>1. Create a plan from a brief: ``` omc ultragoal create-goals --brief-file plan.md ``` Or with explicit stories: ``` omc ultragoal create-goals --brief "ship the migration" \ --goal "Schema::Add new columns" \ --goal "Backfill::Backfill rows in batches" \ --goal "Cutover::Drop old columns and switch reads" ``` The default mode is `aggregate` (one Claude `/goal` covers the run). Pass `--claude-goal-mode per-story` if you want each story to have its own `/goal`. **Multi-repo workspaces / parallel sessions:** when several Claude sessions in the same workspace need to run `/ultragoal` concurrently, pass either `--plan-id <stable-id>` or `--auto-plan-id` so the plan is written to `.omc/ultragoal/plans/{planId}/` instead of the shared single-plan path. Without that flag, two sessions creating goals would clobber each other. `--auto-plan-id` derives `{epochMs}-{slug}` from the brief title. Then thread the same `--plan-id <id>` through every subsequent subcommand in that session. Use `omc ultragoal list-plans` to enumerate available planIds when needed.2. Start (or resume) the next story: ``` omc ultragoal complete-goals ``` This prints a model-facing handoff. The active Claude agent must read it and: - Confirm/Set the active `/goal` condition in this session. - Work the story. - When the story is complete (and for the final story, after the full quality gate), share back a snapshot of the active `/goal` state and call `checkpoint`.3. Checkpoint a story: ``` omc ultragoal checkpoint --goal-id G001-... --status complete \ --evidence "tests/files/PR evidence" \ --claude-goal-json '{"goal":{"objective":"...","status":"active"}}' ``` For the final story, also pass `--quality-gate-json` containing `aiSlopCleaner`, `verification`, and `codeReview` evidence (all clean).4. If the final review is not clean, do NOT mark complete. Record blockers: ``` omc ultragoal record-review-blockers --goal-id G00X-... \ --title "Resolve final code-review blockers" \ --objective "Fix the listed review findings and rerun final gates" \ --evidence "<the review findings>" \ --claude-goal-json '{"goal":{"objective":"...","status":"active"}}' ``` This appends a new blocker story and keeps the Claude `/goal` active.5. Inspect state at any time: ``` omc ultragoal status ```</How_To_Use><Important_Limitations>- The shell cannot invoke or mutate Claude Code `/goal` state. `omc ultragoal` only persists durable artifacts and prints instructions that the active Claude agent reads and acts on in-session.- Snapshots passed via `--claude-goal-json` are model-supplied proof of the active `/goal` state; OMC validates them for textual consistency with the plan's expected objective and ledger event, but it cannot independently observe Claude `/goal` state.- If the Claude `/goal` slash command is renamed or restructured, only the handoff wording needs to change; the reconciliation logic is name-agnostic.</Important_Limitations>
salacoste/oh-my-bmad/tree/main/upstream/omc/skills/ultragoal commit 690b151e7b
Frequently asked questions
Run npx skillmds@latest add salacoste/ultragoal in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.
Durable multi-goal workflow that persists plan/ledger artifacts under .omc/ultragoal and prints Claude /goal handoff text for the active session It is listed under Productivity on SkillMD.
SkillMD's automated safety review verdict for this skill is PASS. Independent scanners report: SkillSpector: PASS, Skill Scanner: PASS. Capability flags: docs only. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.
This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.
Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.
salacoste (@salacoste) published this skill. Their other Agent Skills are listed on their SkillMD profile.