long-term-task-orchestration
Long-Term Task Orchestration Meta-Skill
A framework for designing and generating "long-running task Skills" for AI Coding Agents, with eval and multi-layer review to ensure quality delivery. Characteristics of long-running tasks: large number of files (dozens to tens of thousands), cannot complete in a single session, requires concurrent scheduling, requires cross-session resumption.
Generation Workflow (End-to-End)
Step 1 — Requirements Analysis
Based on the user's description, infer initial answers to the four key elements yourself, then confirm with the user (see full template in references/requirements.md).
Important: Always present your own inferred answers before asking questions. State directly what you can already determine; only use option-style questions for genuinely uncertain elements. Guide the user with directed questions — do not present a blank questionnaire for them to fill in.
Step 2 — Architecture Decisions
Make three key choices based on requirements (decision matrix in references/architecture.md, existing case comparisons in references/examples.md):
- State storage format: TSV / JSON manifest+inputs / JSON single file