How AgentCounsel works
AgentCounsel is a Markdown-native legal skills library. Most tasks are a single skill; larger work uses packs, matter workspaces, playbooks, review panels, and a quality layer. Everything is plain Markdown — no backend, no account, no lock-in.
Skills
Each skill is one workflow with a fixed structure (Purpose, Use When, Required Inputs, Do Not Use When, Legal Safety Rules, Workflow, Output Format, Attorney Verification Checklist). Browse by practice area.
Packs
Practice-area packs bundle every skill in an area — plus the safety rules and commands — into one file you upload to ChatGPT, Claude, or Gemini, with repo-agent variants for Cursor, Codex, and Claude Code. Packs are available for multiple AI tools.
Quality layer
Cross-cutting quality checks review a draft before it is relied upon: source validation, citation integrity, assumption audit, hallucination red-team, privilege/confidentiality review, prose polish, output-format compliance, and the attorney-review gate. See docs/QUALITY_LAYER.md.
Evals and benchmarks
Lightweight, no-API evals check structure, routing, metadata, packs, and candidate-output safety signals. They do not verify legal correctness and do not replace attorney review. See docs/EVALS.md and the coverage report.
Matter workspaces
A matter workspace organizes one matter — parties, facts, sources, deadlines, outputs, and attorney review — as plain Markdown. Start one with py scripts/init_matter_workspace.py from the canonical multi-file template. Choose a workspace over a single skill when the work is multi-step, document-heavy, or deadline-sensitive.
Playbooks
Playbooks are repeatable recipes for recurring task types (NDA review, demand letter, DPA review) — default questions, risk-tolerance settings, required sources, and required quality checks on top of a skill.
Review panels
Review panels run a draft through several supervised review passes (issue spotter, source/citation, client-position, privilege, business-risk) ending with the attorney gatekeeper. The passes are review passes — not autonomous agents and not lawyers.
Platform compatibility
AgentCounsel runs in ChatGPT, Claude, Gemini, Cursor, Codex, Claude Code, and plain Markdown. See platform setup and docs/PLUGIN_COMPATIBILITY.md.
Workflow router
Not sure which surface to use? The workflow router maps "I need to do X" to the right skill or surface, and docs/CHOOSE_YOUR_WORKFLOW.md is a decision guide with worked examples.
Contributing
AgentCounsel is open source under Apache-2.0. See CONTRIBUTING.md for how to add skills, quality checks, evals, playbooks, review panels, and packs — and the safety rules that are not negotiable.