Ultra planning
Skill keithhegit/ultra-orchestration/skills-vnext/ultra-planning
Ultra Orchestration 是一套开源的 AI 工程编排技能体系。它将 OpenSpec 规格主线(Program → Milestone → Change → Slice)与 Ultra-Orchestrator 执行阶段(Intake → Plan → Dispatch → Execute → Review → QA → Deliver → Retro)结合,提升交付稳定性与可审计性。
npx -y skills add keithhegit/ultra-orchestration --skill ultra-planningAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Rigid decision-complete planning for Codex Ultra vNext. Use after intake, brainstorming, or OpenSpec bridge whenever work spans multiple files, roles, phases, or risk boundaries and must be turned into review-ready TaskManifest and WorkPackage artifacts with explicit dependencies, owned paths, acceptance checks, retry assumptions, and serial versus parallel execution rules.
SKILL.md
4.8 KB, as published. Nobody here has run it
Ultra Planning
Use this skill as a rigid workflow gate before controlled dispatch. This skill must remove planning ambiguity rather than summarize it.
Hard Gate
Do not allow controlled dispatch until the plan includes:
- success criteria
- constraints
- non-goals
- owned paths
- acceptance checks
- dependency order
- risk level
- retry assumptions
If any of those are missing or still depend on chat interpretation, the plan is not complete.
When To Use
Planning is required when:
- the work spans multiple files or directories
- multiple phases or roles are involved
- OpenSpec change assets are being turned into execution packages
- review, QA, or risk gates need explicit targets
- parallel execution is being considered
- the implementer would otherwise need to make design or boundary decisions
Planning may be compressed only when:
- the task is truly tiny
- the write scope is narrow and obvious
- no risky command or external mutation is involved
- review and QA targets remain explicit
If you compress planning, record the reason in the orchestration log.
Planning Workflow
- Read the intake, brainstormed design, or OpenSpec bridge artifacts.
- Check whether success criteria, constraints, and scope are already stable.
- If not stable, route back to
ultra-brainstormingor the intake step. - Define the top-level
TaskManifest. - Split the work into one or more
WorkPackageitems. - Mark dependencies, owned paths, parallel groups, and serial boundaries.
- Attach acceptance checks, risk level, and retry assumptions.
- Record assumptions, blockers, and re-plan triggers.
- Run the planning self-review before declaring the plan complete.
Planning Standard
Produce one TaskManifest plus one or more WorkPackage items.
The plan is not complete until it defines:
- goal
- context
- success criteria
- constraints
- non-goals
- dependencies
owned_paths- expected outputs
- risk level
- acceptance checks
- retry budget or retry assumptions
DAG And Lock Rules
For each work package, identify:
- upstream dependencies
- whether the task is parallel-safe
- which paths are read-only context
- which paths are write-owned
Only mark work as parallel-safe when both conditions hold:
- all upstream dependencies are satisfiable without new design work
owned_pathsdo not intersect with another active write package
Completion Standard
Planning is complete only when:
- a new implementer could start from the plan without inventing key decisions
- review knows what to inspect
- QA knows what behavior must be verified
- write ownership is explicit
- serial versus parallel execution is explicit
- blockers and re-plan triggers are named
If the plan still requires the implementer to decide boundaries, acceptance, or parallelism, the plan is incomplete.
Planning Self-Review
Before returning the plan, check:
- do any
owned_pathsoverlap unexpectedly - are acceptance checks concrete and verifiable
- do reviewer and QA receive explicit inspection targets
- is any work package write scope too broad
- do dependencies still rely on chat memory rather than artifacts
- does any package mix planning, implementation, and review responsibility
- would a failure in one package force a re-plan that is not yet documented
Failure Handling
- If success criteria, scope, or constraints are still unstable, route back to
ultra-brainstorming. - If OpenSpec assets are incomplete or contradictory, route back to
openspec-ultra-bridge-v2or spec preparation. - If planning reveals that one request actually contains several independent initiatives, decompose the work and plan only the first bounded unit.
Role Discipline
- planner defines work, but does not silently become the implementer
- reviewer tasks must not be mixed into implementation packages
- integrator tasks should package accepted outputs, not redo worker code
Output Template
Return in this order:
- one
TaskManifest - ordered
WorkPackageitems - parallel groups or serial sequence notes
- assumptions
- blockers
- re-plan triggers
Examples
- Read planning-workflow for the rigid flow.
- Read planning-anti-patterns for what makes a plan unsafe or incomplete.
- Read work-package-examples for one normal multi-file feature and one OpenSpec change route.
- Read planning-checklist before finalizing.