Add Legal Verification Pack (4 assistant workflows)

[mazurwieslaw2022-cmd]

Add Legal Verification Pack (4 assistant workflows)

What

A new assistant workflow pack, legal-verification-pack, with four jurisdiction-neutral workflows for stress-testing a finished legal deliverable before it goes out. They form a cost gradient, cheapest first:

• judicial-first-impression - a neutral cold-reader assessment: how the submission lands on a judge reading it for the first time, under time pressure, with no prior context. Seven-part reception report, including a calibrated confidence level and the list of gaps that would need closing to persuade.
• opposing-counsel-attack - a single-pass attack from the perspective of counsel for the other side, told to find every way to beat the argument. Six sections, from the core theory of attack to what the submission is trying to hide. Includes procedural angles: burden of proof, time limits, admissibility, scope of appeal.
• adversarial-legal-review - the top rung, cost-gated for high-stakes work: a full debate in four roles (builder, attacker, synthesizer, verifier) with a ten-point final check and a bounded revision loop that can never ship a version worse than the best one seen.
• output-scoring - a two-layer scorecard that keeps mechanically checkable facts (citations exist, dockets real, no finding dropped) apart from rubric-based judgement (five dimensions, 1-5), ending in a send / revise / verify decision.

Every workflow assesses or attacks an existing text; none drafts, files, or sends - those acts stay with the lawyer. All four state an explicit negative scope and use a [VERIFY] tag for any authority not checked in-session, so a fabricated reference cannot pass silently.

Where it comes from

The workflows are adapted from the open-source collection at awesome-matematic-skills-en and rewritten for this repository: harness-neutral wording, no CLI commands, Mike frontmatter, add-on availability.

Third-party material, with attribution kept inside each workflow: opposing-counsel-attack and judicial-first-impression adapt skills by Larissa Meredith-Flister (lawvable/awesome-legal-skills, Apache-2.0 as declared in the author's frontmatter) - the role framing and section structures are retained, the jurisdiction-neutral framing and governance boundaries are new. adversarial-legal-review follows the debate pattern of AnttiHero/lavern (Apache-2.0) with roles and checks written from scratch, plus loop-control concepts adapted clean-room from gregmos/memoforge (MIT). output-scoring bases its two-layer method on DISC-Law-Eval (FudanDISC/DISC-LawLLM, Apache-2.0) with the rubric and thresholds written from scratch.

Overlap with existing workflows

The existing review workflows (nda-review, msa-review, dpa-review and peers) review a CONTRACT for its terms. This pack reviews an ARGUMENT for its weaknesses - the deliverable your own side produced, before it goes to the client, the court, or the other side. The two sets compose: a contract review produces a memo; this pack stress-tests the memo. Within the pack the four workflows are deliberately non-overlapping (neutral reception / hostile single-pass / full debate / scoring decision), and each one's SKILL.md states when to use which.

How it was validated

workflow-schema/validate-workflows.py passes with the pack in place: 138 workflows, 16 column files, 5 packs. Checked by hand for what the validator does not cover: no CLI blocks, no harness-specific tool names, no em dashes in descriptions.

🤖 Generated with Claude Code