Skill v1.0.1
currentLLM-judged scan95/100+1 new
version: "1.0.1" name: hiveward-skill-decomposer description: Use when turning a skill package, standalone Markdown skill, SKILL.md, skill folder, or script-backed skill into a HiveWard blueprint proposal.
HiveWard Skill Decomposer
Purpose
Use this skill to turn supplied skill material into a governed HiveWard blueprint proposal. A skill can be a full package, a standalone Markdown skill, pasted Markdown, a local path, a URL, or repository material that resolves to one of those forms.
The decomposer does not install arbitrary user skills as global HiveWard skills. It analyzes the material, builds Skill IR first, maps that IR into a blueprint proposal, and submits only through HiveWard inbox approval when the user explicitly asks.
Required Flow
- Locate or receive the skill source.
- Classify source completeness as
full_package,markdown_only,partial_package, orunknown. - Inventory the package root when one exists.
- Read the primary Markdown entry point.
- Inspect supporting folders when available.
- Build Skill IR.
- Validate Skill IR.
- Map Skill IR to a blueprint proposal.
- Validate the proposal against the current HiveWard blueprint and inbox contracts.
- Explain unresolved assumptions.
- Submit to inbox only when the user asks for formal approval.
Core Rules
- Treat a skill as a package, not only
SKILL.md. - If a user identifies one Markdown file as the whole skill, set source completeness to
markdown_onlyand do not invent missing folders. - If only
SKILL.mdcontent is available for a package-like skill, set source completeness topartial_packageand mark sibling inventory unresolved. - Do not execute scripts by default. Inspect scripts statically as controlled assets.
- Build Skill IR before blueprint nodes. The IR is the contract.
- Prefer a smaller correct blueprint over a large speculative one.
- Split only on real work boundaries: independent inputs/outputs, distinct tools or permissions, meaningful validation, safe parallelism, retry/failure branches, script side effects, or decision points.
- Record phase difficulty, model class, desired thinking effort, and parallelism hints in Skill IR.
- Do not claim per-node thinking effort is enforced until HiveWard runtime schema supports it.
- Preserve HiveWard platform boundaries: deterministic routing, lifecycle, persistence, permissions, and artifact publication belong to HiveWard; model judgment and content belong in prompts, role contracts, schemas, and examples.
- Do not claim a blueprint changed until inbox approval/import confirms it.
Blueprint Mapping
Use current HiveWard node types: agent, manager, manager_slot, condition, summary, note, group, and loop.
- Role skills default to notes/groups plus an operating brief unless the user asks for an executable workflow.
- Simple process skills should stay small.
- Multi-phase or script-backed skills can use a manager with slots and a final summary.
- Scripts remain files or path references; they do not become a new node type.
- Script-aware agents must state script path, working directory, command template, inputs, outputs, permissions, validation, approval needs, and side effects.
- Runtime ids for blueprint nodes are
codex,claude,openclaw,hermes,google,cursor, andopencode. Prefer the operator priority order Codex, Claude Code, OpenClaw, Hermes, then other CLI harnesses unless the source skill or user requires a specific harness. - Use node-level
runtimeId; never placeruntimeIdinsideconfig. Useclaudefor Claude Code blueprint nodes andclaudeCodeonly when discussing harness/status APIs. - For OpenClaw
agentand runnablemanagernodes, includeconfig.openclawAgentIdonly when the target Agent is known or intentionally selected, and includeconfig.modelIdonly when the OpenClaw model choice is intentional. Non-OpenClaw nodes must not carryopenclawAgentIdor OpenClawsendconfig. Hermes nodes may useprofileId. - Manager modes are explicit field pairs: sequential is
lifecycleMode: "none"plusdispatchMode: "sequential", self-dispatch islifecycleMode: "none"plusdispatchMode: "self_dispatch", and self-iteration islifecycleMode: "self_iteration"plusdispatchMode: "self_dispatch". - Manager slots are containers, not ordinary chain nodes. Child nodes inside a slot must set
parentIdto the slot id. Do not connect manager slots to each other as a sequence. - Standard Manager/slot handles are
manager-out-N->manager-slot-in,manager-slot-out->manager-in-N,manager-slot-inner-outfrom slot to first child, andmanager-slot-inner-infrom last child back to slot. - Use
resultRole: "final"for the intended deliverable,resultRole: "ignore"for internal Manager-slot workers, and omitted/autofor normal terminal outputs. - Use
summary.config.mode: "structured_merge"for deterministic aggregation andharness_summaryonly when a runtime summary agent is genuinely needed. - A formal proposal must be returned as ordinary assistant text plus a complete blueprint package proposal; its
blueprintPackage.schemamust behiveward.blueprint-package/v1.
References
Load only the reference files needed for the current decomposition:
references/skill-package-model.mdreferences/decomposition-rules.mdreferences/blueprint-mapping-rules.mdreferences/script-handling-rules.mdreferences/execution-planning-rules.mdreferences/skill-ir-schema.md