Skills for legal work and building — public offerings for the legal tech community.
Rapid, IP-focused triage of an AI tool's Terms of Service before corporate adoption. Built for standard, non-negotiable click-through terms where the decision is up-or-down: it zeroes in on the two questions that can be existential — can the provider train on your data, and do you own the outputs — and deliberately ignores the general terms (liability caps, indemnities) that create risks you can insure against or live with.
Returns a one-line verdict, an employee-facing findings list, and a section-cited, quote-backed addendum for the reviewing attorney. Model-neutral: runs as an agent skill (Claude Code, etc.) or pastes directly into a Claude Project / ChatGPT custom GPT. Ships with a runnable eval set.
Intended workflow: require requesters to run this on the terms before submitting a legal-review ticket and to include the output in the ticket. Bad terms get caught before legal ever sees them; good ones arrive pre-digested with the key clauses located and quoted. Triage, not clearance — every verdict still routes to a human attorney.
A US-only trademark clearance workflow: structured seven-question intake (hard gate — no analysis until answered), live USPTO federal search covering registrations and pending applications with dead-mark follow-up, common-law digital-footprint sweep (web, app stores, domains, social), Abercrombie strength plus a full refusal-ground screen (§2(e)(1)–(4), §2(c), §2(a), ornamentation/failure to function), an explicit priority determination, 13-factor DuPont analysis, a separate dilution screen, and a source-tagged report built for attorney review.
Two rules make it safe for legal workflows: it never fabricates registration data — only records read from a live USPTO/TSDR source this session count as verified, everything else is tagged [Requires Manual Verification], and the same doubt applies to statutory and TMEP pincites — and it calibrates its searches, never treating a zero-result query as a clearance until proven syntax has returned a known record. Two screens run outside the confusion analysis because DuPont is blind to them: priority (who was first) and dilution (a near-copy of a famous mark on unrelated goods scores well on DuPont and no examiner will ever raise it). State registries and non-US registers are deliberately excluded and the report says so. A knockout mode gives a fast kill/proceed screen on a reduced intake gate, and a CANDIDATES list mode compares several names at once. Ships with a runnable eval set including fabrication-resistance, jurisdiction-boundary, and live-dispute tests.
Intended workflow: run it on every candidate name before anyone falls in love with one; the report routes to attorney review with the key records located, quoted, and pre-verified.
Download a PandaDoc document as a PDF from its public view link (app.pandadoc.com/document/v2?token=...), including when the sender has disabled the recipient's download button. Produces two files: PandaDoc's own byte-exact stored source PDF, and a signed reconstruction with the signature images and field values stamped in at the positions the viewer renders them. Built on Playwright browser tools; the link's token is the user's own document access.
Turn a goal into self-contained GitHub issues, then grind through them AFK with a deterministic multi-agent loop: fresh clean-context coder per ticket, independent adversarial reviewer who re-runs all verification, committer that lands with evidence — and blocked tickets are parked (work stashed, findings posted to the issue, triage label applied) so an overnight queue never dies at the first stubborn ticket. Model-agnostic: roles are capability tiers and effort levels, not model names. Requires a runtime with a Workflow-style orchestration tool (built for Claude Code).
Not a skill: a provisioned, verified headless OpenSCAD toolchain for agent-driven mechanical design, plus a survey of existing agent CAD skills. install.sh sets up the toolchain and libraries, verify.sh is an acceptance test that must catch a motion clash without inventing one, and render.sh produces six views and a contact sheet to read. Includes a worked project (projects/pill_organizer_10bay). Use it by cloning the repo; it is not offered as a plugin.
As Claude Code plugins, one per skill:
/plugin marketplace add LegalMarc/skillz
/plugin install tm-clearance@legalmarc-skillz
Replace tm-clearance with any of ai-tos-review, pandadoc-pdf-download, tm-clearance, or workflow-loop.
Manually: clone this repo and copy or symlink the skill directory you want into ~/.claude/skills/, e.g. ln -s "$PWD/tm-clearance" ~/.claude/skills/tm-clearance. The skills are plain SKILL.md folders, so they also work in any agent runtime that reads that format.
ai-tos-review and tm-clearance each ship an original-prompt.md — the verbatim chat prompt the skill was distilled from, kept as a historical artifact. If you're converting your own working prompts into skills, the before/after pairs show exactly what conversion adds: the domain expertise carries over nearly untouched, while verification gates, deterministic rubrics, self-checks, and eval sets grow around it.
MIT — see LICENSE.