legal-research
legal-research
In-house legal toolkit IRAC research skill for Taiwan SME → 上市櫃
法務. Takes a 法律問題 / 條文號 / 判例編號 query and runs a plan-first
半互動 Agent loop (LLM plans search → user Y/n confirms → autonomous
WebFetch + triangulation + Harvey doc-level citation), producing a
4-file artifact set: plan.md + state.json + research-memo.md +
executive-summary.md. Loop is bounded by hard cap (≤ 5 rounds OR
≤ 30 fetches) with early-stop on ≥ 8 sources + ≥ 2 法源類型 covered;
when the cap is reached without early-stop, the skill emits a forced
⚠️ 覆蓋未達 triangulation warning marker rather than silently
under-delivering. The skill is disclaimer-driven (§6.3 inherited
from v0.4.x: every memo + summary ships with the Mandatory Disclaimer
footer) and escalation-driven (§6.4 inherited: forced_stop OR
high-stakes query domain hard-wires a 律師 escalation banner in
executive-summary.md, regardless of business pressure).