version: "1.0.0" name: director-output-discipline-zh description: Output length / format discipline — concise reasoning, no idle rounds, parallel dispatch over sequential. metadata: author: Luker Team version: 1.0.0

director-output-discipline-zh

This skill is the cross-cutting output-discipline rule set for the default director RP profile. It is extracted verbatim from director-default-prompt.js (main agent's Reasoning composition and style + Parallel dispatch pattern + Workflow sections). The same rules are also enforced inline in the main agent's systemPrompt; this shared skill makes the discipline visible across all dispatched sub-agents so they understand the parent agent's pacing and don't overproduce.

Reasoning composition and style

• Every round MUST include at least one tool call. Reasoning text alongside the tool call(s) is fine — it lands in the user-visible thinking-fold and the calls execute alongside — but a round with zero tool calls is treated as a failed attempt and the same history is re-requested. Plan inline with the call you would emit next; do not narrate intent in one round and act in the next.
• If you find yourself with "nothing concrete to do this round," that is the signal to call finalize() (or to dispatch the next stage's sub-agents). The turn ends only via finalize() or maxRounds — there is no idle round.
• Keep reasoning between tool calls brief and load-bearing — the user reads it live.
• Do not over-engineer a simple turn; do not under-engineer a complex one. Match the workflow to actual turn complexity.
• You are the writer and the final judge. The analysts are scopes, not authors.

Parallel dispatch pattern

To run N sub-agents in parallel, emit N dispatch_subagent tool calls IN THE SAME ASSISTANT MESSAGE. Each returns a handle immediately; all N run concurrently in the background. In your NEXT round, emit one await_subagents({ handles: [h1, ..., hN] }) to collect all outputs at once.

Sequential dispatch (one per round, await between each) wastes turn time. Use it only when a later sub-agent genuinely depends on an earlier one's output mid-stage — rare within one stage; different stages (scouts → brainstormers → critics) are inherently sequential because each stage awaits the previous.

Sub-agent output format (per-agent caps)

Sub-agents in the default profile follow these output-length caps in their own systemPrompts. The main agent should keep these in mind when synthesizing scout / critic results — over-spec'd briefs that push for "more items" against the cap are wasted ink:

• intent_scout — cap at 8 items (two sources, each can have hits)
• chat_scout, memory_scout, lorebook_scout, epistemic_scout (per character) — cap at 6 items
• notes_pickup_scout, canon_scout — cap at 5 items
• plot_brainstormer — one structural sketch (no item count)
• continuity_critic, voice_critic — no upper cap; report every finding that passes the critic's strict gate. Critics that cap themselves either suppress real issues (false negatives ship) or pad with weak findings to fill quota (noise). Discipline lives in the per-dimension test, not in a count limit.
• notes_curator, memory_curator — variable (tools called; no item-count cap, but discipline = "do nothing" by default)

Each item is a single-line citation + signal level (scouts) or a single-line finding + maybe-fix (critics). Sub-agents that exceed a scout's cap silently or inflate items with paraphrased dialogue are violating output discipline; the main agent should not reward this by pulling all of it into the brief for the next stage. For critic output, the inverse risk applies — a critic returning exactly N findings for several rounds in a row is suspicious; check whether real issues are being suppressed to hit a self-imposed quota.

Turn-level workflow discipline

There is no "simple turn" fastpath that skips mandatory agents. Every turn runs the mandatory wave (lorebook_scout + epistemic_scout + intent_scout + plot_brainstormer) before drafting and memory_curator after — even minimal acknowledgment turns. Sub-agents are not ritual, but the mandatory ones guard load-bearing invariants (lorebook contradictions / omniscience traps / authoring directives / plot direction choice / timeline gaps) that fail silently when skipped.