Opinionated skill packs — pstack shape for orchestrators
Opinionated skill packs — pstack shape for orchestrators
Section titled “Opinionated skill packs — pstack shape for orchestrators”Related: 2026-09-25 Cursor Project and pstack · Cloud agent orchestrator
This deepens skill-pack shape for product keypoint 5; it does not rewrite the Cursor Projects / pstack overview note.
pstack’s layered shape
Section titled “pstack’s layered shape”pstack is not a flat slash-command dump. Layers:
- Router (
/poteto-mode) — maps a natural-language goal + finish/verification condition to a playbook. - Playbooks — task protocols (bugfix, feature, refactor, orchestrate, overnight, ship).
- Leaf skills — reusable actions (
how,architect,tdd,unslop,show-me-your-work, …). - Principles — cross-cutting constraints (“prove it works,” “sequence verifiable units”); indexed by the router, loaded when needed.
- Scripts / CLIs — durable orchestration helpers (audit trails, worktrees, parallel workers).
Opinionation sits at routing + protocol. Leaf skills stay reusable instructions. That is the right split for a DIY orchestrator pack.
Sources: pstack guide · poteto-mode · orchestrate playbook · overnight contract.
Analogous packs
Section titled “Analogous packs”| Pack | Shape note | Links |
|---|---|---|
| Agent Skills (Anthropic open standard) | Dir + SKILL.md; metadata first, instructions on activation, resources progressive | spec · anthropics/skills |
| Cursor Skills | Discovery, monorepo paths, explicit-only, custom modes, plugins | docs |
| Continue | SKILL.md + older rules vs skills; typed core protocol | rules · protocol |
| OpenCode / Claude Code | Agent Skills + permissions (allow/deny/ask); slash→skills; plugin marketplaces | OpenCode · Claude · official plugins |
Recommended pack layout (keypoint 5)
Section titled “Recommended pack layout (keypoint 5)”pack/ manifest.json router/SKILL.md principles/*.md playbooks/ bugfix.md feature.md investigation.md orchestrate.md skills/ understand/SKILL.md design/SKILL.md verify/SKILL.md summarize/SKILL.md scripts/ adapters/ cursor.md claude.md continue.md opencode.md pi.md evals/Each skill should declare: frontmatter name/description, triggers + non-triggers, inputs/outputs, allowed tools + side effects, short procedure, checkable exit condition, optional progressive scripts/refs.
Design rule: Opinionation belongs in defaults, invariants, and verification gates. Flexibility belongs in adapters, leaf skills, model choice, and user overrides. Principles should be named, individually disableable, and testable. Add evals for “should trigger,” “should not,” and outcome quality.
Platform adapters translate abstract actions (spawn_worker, verify_artifact, read_events) into Pi / Cursor / Claude / Continue / OpenCode tools — do not bake one host’s style into every leaf skill.