Luban Complex Task Execution Coach

Problem

Users often provide only a vague request, such as “organize these materials and give me the conclusion.” If an agent starts immediately, it can silently assume context, over-engineer, modify out-of-scope items, or treat internal logs as the deliverable. Luban Skills is positioned as a delivery-focused task execution coach: first define what the final result should look like, then decide whether to call Hermes, ACP, scripts, or read-only commands.

How It Works

The workflow has two layers. For ambiguous requests, it follows a six-step process: classify the task as writing, organization, analysis, planning, or execution support; define an acceptable deliverable; extract file, format, prohibition, and confirmation constraints; build a minimal closed loop; proceed in phases; and validate against whether the user can directly use the result. When Hermes is needed, it uses an Intent / ReAct / Reflect loop: define goal, constraints, risks, and first safe action; iterate with small, evidence-backed steps; and review at checkpoints whether to stop or ask for confirmation.

The skill also emphasizes operational guardrails: long ACP tasks should be split, bound with TIME_BUDGET, checkpointed, retried with limits, and reported with progress events; Feishu webhooks can notify start, progress, need_confirm, timeout, and related lifecycle events; guard and allowlists constrain tool calls. For multi-image tasks, the main goal is persisted image files, not only prompts or storyboards.

Boundaries

It is suited for writing, material organization, table analysis, project planning, and read-only command assistance. It should not be used for high-risk deletions, production changes, medical/legal/financial conclusions, or strong conclusions without reliable evidence. When the user's goal is still unclear, the workflow should narrow scope or request confirmation instead of pretending completion.