Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add eigenweltlabs/legalwork --skill benchmark-taskgit clone --depth 1 https://github.com/eigenweltlabs/legalworkWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/eigenweltlabs/legalwork/benchmark-task)<a href="https://agentmods.dev/skills/eigenweltlabs/legalwork/benchmark-task"><img src="https://agentmods.dev/badge/skills/eigenweltlabs/legalwork/benchmark-task.svg" alt="Measured on agentmods" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to medium
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- medium Rogue Agent · line 4 Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.Fix: Remove any persistence mechanisms (cron jobs, startup scripts, state files). Skills should not maintain state across sessions without explicit user consent.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00076 | $0.00819 |
| Opus 5 | $0.00038 | $0.00409 |
| Sonnet 5 | $0.00015 | $0.00164 |
| Haiku 4.5 | $0.00008 | $0.00082 |
Grade A, and why
benchmark-task scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 58 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Create a benchmark task
The LegalWork benchmark compares how connected AI models perform on realistic legal
tasks. Each task has instructions, expected deliverables, and a rubric of independently
judged pass/fail criteria. This skill turns a user's description (or an open matter)
into a well-formed task and saves it to the benchmark's Tasks table with the
benchmark_create_task tool.
What a good task looks like (the Harvey Legal Agent Benchmark shape)
- title — a short, specific name, e.g. "Review PO1 against the 2021 ICC Rules".
- workType — one of
analyze,draft,review,research. - tags — practice-area / vertical tags; the first tag is treated as the vertical
(e.g.
["Arbitration", "ICC Rules"]). - instructions — directional instructions to the model, in the second person, that
name the expected output file(s). Example: "Review the attached procedural order against
the 2021 ICC Rules and produce
compliance-report.docxlisting each deviation …". - documents (
documentPaths) — INPUT files the model must read, passed as file paths (workspace-relative or absolute), e.g.["documents/procedural-order.docx"]. These are attached to the task and staged for every run. Do not confuse them with deliverables (the model's OUTPUT). If the user refers to files in the workspace/matter, pass their paths here; you may also write a file first (e.g. save pasted text) and then pass it. - deliverables — the output filenames the model must produce, e.g.
["compliance-report.docx"]. Optional but recommended; criteria reference them. - criteria — the rubric. Each criterion is judged independently, so make every one
atomic and objectively verifiable. Write each
matchCriteriaas an explicit rule: "PASS if <specific, checkable condition>. FAIL if is not met." Anchor it to concrete facts (article numbers, dates, names, amounts) wherever you can.
How to do it
- Gather the task. If the user pointed you at a document or matter, base the task on
it and attach the relevant source files via
documentPaths. Ask only for what's genuinely missing; infer sensible values otherwise. - Draft the fields above. Aim for 5–15 sharp criteria unless the user asks for more — each testing one distinct, checkable thing. Don't invent criteria that can't be judged objectively from the deliverable.
- Call
benchmark_create_taskwith the fields, includingdocumentPathsfor any input files the model needs to read. - Confirm what was created (title, number of criteria, and any documents attached) and tell the user it's now in the benchmark's Tasks tab, where they can edit it, add or remove documents, or include it in a run.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 58 lines · 76 tokens per session scan A 16f116c700b7
benchmark-task is a skill published in the GitHub repository eigenweltlabs/legalwork (100 stars, last pushed today), licensed MIT. It adds 76 tokens to every session and 819 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
policy-creation
Guide for adding new enterprise policies to Chromium. Complete guide for policy definitions, pref mapping, and writing tests.
ring:adding-multi-tenancy
Adding database-per-tenant isolation into a Go service end-to-end via an 11-gate cycle: detects the stack, audits compliance, then dispatches backend agents to implement tenantId-from-JWT routing through the lib-commons v5 dispatch layer (config, middleware, repositories, metrics, tests) and runs reviewers. Use when…
dt-sec-insights
Query and analyze Dynatrace security data in security.events with DQL: vulnerabilities, threat detections, compliance posture, and scan coverage. Covers Dynatrace-native Runtime Vulnerability Analytics (RVA — CVEs, reachability, exposure, exploit), Runtime Application Protection (RAP), Automated Detections, and…
ring:testing-skills-with-subagents
Hardening a skill so it resists rationalization under pressure: run combined-pressure scenarios on a subagent without the skill (RED), capture excuses verbatim, write/edit the skill (GREEN), then plug loopholes (REFACTOR) until compliance holds. Use when deploying or editing a discipline-enforcing skill that an agent…
lians-integrate
Wire Lians memory into an existing agent codebase, test-first and minimal-diff. Use when the user asks to add persistent/compliance memory to their agent, integrate Lians, or replace a vector store with bitemporal memory.
accessibility-testing
Test for WCAG 2.2 AA compliance with axe-core + Playwright, keyboard navigation audits, screen reader testing, ARIA pattern validation, and legal compliance mapping (ADA, EAA, Section 508). Automated tools catch 30-40% of issues — this skill covers automated and manual testing together. Use when: "accessibility,"…