AgentCounsel

An open, Markdown-native library of legal skills for AI agents — and the legal professionals who supervise them.

AgentCounsel is a collection of standalone legal skills. Each skill is a structured workflow that helps an AI agent — or a lawyer — produce draft legal work product for attorney review. The skills supply discipline and structure: how to scope a task, what inputs to gather, how to organize the output, and what a supervising attorney must verify before anything is relied upon.

AgentCounsel does not provide legal advice. It produces draft work product intended for review and adoption by a qualified, licensed legal professional. Using it does not create an attorney–client relationship.

New here? Start with QUICKSTART.md — it walks through your first skill run in about five minutes.

Who AgentCounsel is for

• Practicing lawyers and in-house legal teams who want AI assistance that is consistent, reviewable, and built around attorney oversight instead of around a vendor's product.
• Legal operations and knowledge-management teams standardizing how their organization uses AI for routine legal work — intake, triage, first-pass review, and drafting.
• Law students and new lawyers who want to see how experienced practitioners structure a contract review, a chronology, a research memo, or a risk assessment.
• AI builders and engineers assembling legal agents or assistants who need well-scoped, safety-first workflow definitions they can drop into any model or framework.

Whoever you are, the assumption is the same: a licensed attorney reviews the output before it is used.

What it is — and is not

It is:

• A library of legal workflows and output structures.
• A way to make AI-assisted legal drafting more consistent, reviewable, and safe.
• Useful piecemeal (one Markdown file) or as a whole.
• Plain Markdown — no build system, no runtime, no account, no lock-in.

It is not:

• Legal advice, or a substitute for a licensed attorney.
• A source of jurisdiction-specific legal rules. Skills give you process and structure, not the law of any jurisdiction.
• A finished-document generator. Every output is a draft that an attorney must review.
• Built around any single AI product. It is a portable library with thin adapters for many environments.

Quick start

QUICKSTART.md is the one place to start — pick your platform, install a practice-area pack, and run your first skill in about five minutes. Sample outputs are in examples/.

You do not need to read core/ to get started: every pack and adapter loads those operating rules for the AI automatically. Read core/ yourself only if you are using a raw SKILL.md file on its own, or you want to understand the rules your AI is following (a worthwhile ten minutes, but not a prerequisite).

Practice areas

AgentCounsel has 212 skills: 177 across 20 practice areas, plus 35 cross-cutting skills in three supporting groups (Setup, Legal Operations, Legal Methodology).

Three cross-cutting skill groups support work in every practice area:

• Setup (20 skills) — cold-start interviews that configure AgentCounsel for a practice group, plus a matter-workspace builder.
• Legal Methodology (11 skills) — source validation, citation integrity, assumption audit, hallucination red-team, attorney-review gate, privilege/confidentiality review, legal prose polish, output-format compliance, statutory interpretation, risk assessment, and red-team verification.
• Legal Operations (4 skills) — legal intake triage, templated legal responses, meeting briefings, and signature-routing checks.

The four-tier taxonomy that classifies these areas and groups — and the model for expanding it — is recorded in docs/PRACTICE_AREAS.md.

See SKILLS_INDEX.md for the full catalog and WORKFLOW_ROUTER.md to route from a task to a skill.

Choose your workflow

Most tasks are one skill. When the work is larger, AgentCounsel gives you several surfaces:

Full decision guide with worked examples: docs/CHOOSE_YOUR_WORKFLOW.md. Every surface ends in attorney review.

How it is organized

core/               Operating rules every skill inherits (read these first).
skills/             The canonical skill library, grouped by practice area.
examples/           Illustrative sample outputs, with fictional facts.
practice-profiles/  Per-practice-area configuration profiles for a legal team.
matter-workspaces/  Scaffolds for organizing one matter — single-file templates
                    plus the canonical multi-file _template/.
matter-packs/       Workflow bundles — recommended skill sequences per matter type.
playbooks/          Repeatable recipes for recurring task types (NDA review, etc.).
review-panels/      Supervised multi-pass review workflows for a draft.
overlays/           Industry and sector overlays that tune skills for a context.
adapters/           Thin integration files for specific environments.
connectors/         External-source verification integrations (e.g., CourtListener).
metadata/           Generated machine-readable skill index (index.json).
docs/               The metadata standard, the safety model, the FAQ, and the practice-area registry.
scripts/            Standard-library Python helpers (validation, index, packs).
dist/               Generated platform install packs (gitignored; built by
                    scripts/build_platform_packs.py).
SKILLS_INDEX.md     Full catalog of every skill.
WORKFLOW_ROUTER.md  "I need to do X" -> which skill to use.
COMMANDS.md         Slash-style command shorthands mapped to skills.

The skills/ directory is the canonical source of truth. Everything else points back to it.

Anatomy of a skill

Each skill lives in its own folder with a SKILL.md file and optional templates/:

skills/contracts/nda-review/
  SKILL.md                     The workflow.
  templates/nda-risk-table.md  A copyable, attorney-review-ready template.

Every SKILL.md follows the same structure — Purpose, Use When, Required Inputs, Do Not Use When, Legal Safety Rules, Workflow, Output Format, and an Attorney Verification Checklist — and carries standardized, agent-readable YAML frontmatter so tools and agents can discover and route skills. See docs/SKILL_METADATA_STANDARD.md.

Ways to use AgentCounsel

Recommended: install a pre-built practice-area pack

A platform pack is one file that bundles the global safety rules, the practice profile, the slash commands, and every skill in a practice area. Upload it to your AI tool's project or workspace and you can start work immediately — no copying files one at a time.

Pre-built packs are hosted on the deployed Pages site and rebuilt on every push to main:

• Browse the packs page — direct downloads per practice area for ChatGPT (.md), Claude (.zip), and Gemini (.zip), plus AGENTS.md / CLAUDE.md for repo agents.
• The packs page lists every practice area with its skill count and a "How to use" block per platform.

For finer-grained control, you can also...

AgentCounsel is plain Markdown, so it works anywhere an agent or person can read a file. The library-as-files paths below give you more control than a consolidated pack.

Adapters are intentionally thin: they tell an environment how to find and use the canonical library; they do not duplicate it. The one exception is the Claude Code plugin bundle, which carries generated copies of a curated set of skills — regenerate it with python scripts/sync_plugin_skills.py (see PLUGIN_SYNC.md).

Installing or copying skills

• A practice area (recommended): download a pre-built pack from the packs page — one file per practice area for ChatGPT, Claude, and Gemini. Or run python scripts/build_platform_packs.py locally to produce the same packs in dist/.
• One skill: copy a single SKILL.md (and its templates/, if any) into your assistant. Nothing else is required.
• The whole library: clone the repository, or vendor skills/ and core/ into your own project.

There is no installer, no package to publish, and no runtime dependency. A skill is just a file.

Safety model

Every skill inherits the operating rules in core/:

• Draft legal work product only. Attorney review is always required before output is relied upon.
• Never invent legal authority, citations, quotations, facts, or deadlines.
• Separate facts, assumptions, law, analysis, strategy, and verification items.
• Identify jurisdiction, governing law, posture, and the relevant date — or flag them as unknown.
• Preserve confidentiality and privilege.
• Flag uncertainty with visible placeholders instead of guessing.

The full safety model — the threat model, what AgentCounsel deliberately does not do, and how to use it safely with confidential material — is in docs/SAFETY_MODEL.md. Security reporting and guidance are in SECURITY.md.

A firm's risk or compliance team assessing AgentCounsel for adoption should start with docs/AI_GOVERNANCE_DISCLOSURE.md, which maps these controls to common AI-governance framework questions.

Configuring AgentCounsel for a team

AgentCounsel runs out-of-the-box without configuration. To improve output quality for a practice group adopting the library across multiple matters, you can populate a practice-group profile at practice-profiles/<area>.md. The profile gives ~20 skills access to the group's standing positions, escalation thresholds, and output preferences.

Eval coverage is tracked through lightweight, no-API checks in evals/ and reports/eval-coverage.md. These evals test structure, routing, metadata, packs, quality checks, and candidate-output safety signals; they do not verify legal correctness or replace attorney review.

See CONFIGURING.md for the full configuration guide. The short version:

1. Run the cold-start interview for the practice area (skills/setup/<area>-cold-start-interview/SKILL.md) with a supervising attorney, or edit the profile template directly.
2. Load the populated profile into your AI workspace alongside any skill that consumes it (see practice-profiles/README.md for the list).
3. Skills that consume the profile will benchmark against its Standard Positions and Escalation Thresholds; skills that do not consume it are unaffected.

Configuration is opt-in. Skipping it costs you benchmarking-against-standing-positions but does not degrade any other skill's output.

Two optional layers let a legal team adapt the library without changing any skill:

• practice-profiles/ — one profile per practice area capturing the team's jurisdictions, escalation thresholds, standard positions, review requirements, and prohibited assumptions. The Setup cold-start interview skills fill a profile in.
• matter-workspaces/ — scaffolds for organizing one specific matter: parties, documents, deadlines (always flagged for attorney verification), open items, sources, and an index of the draft work product produced by skills. Single-file templates plus a canonical multi-file template at matter-workspaces/_template/, which python scripts/init_matter_workspace.py instantiates. See docs/MATTER_WORKSPACES.md.

Both are plain Markdown and add no backend, runtime, or vendor dependency.

Examples

The examples/ directory contains illustrative sample outputs — including primary skill outputs and quality-layer passes such as source validation, prose polish, hallucination red-team, and attorney-review gating. Every fact in them is fictional. They show the shape and quality of a skill's deliverable; they are not legal advice and not templates for a real matter. If you have not run a skill end-to-end yet, docs/tutorials/practice-matter.md is a guided, fictional first-run walkthrough.

Validation and continuous integration

The repository ships a lightweight validation script — Python standard library only, no dependencies — that checks every skill's structure, frontmatter, links, and safety framing:

python scripts/validate_repo.py

GitHub Actions runs the same checks on every pull request and on pushes to main (see .github/workflows/validate.yml):

python scripts/sync_plugin_skills.py --check   # plugin bundle is in sync
python scripts/validate_repo.py                # full repository validation

Other standard-library helpers in scripts/ build the machine-readable skill index (build_skill_index.py), the per-platform install packs (build_platform_packs.py), the matter-workspace initializer (init_matter_workspace.py), and the browsable static catalog under site/. Every script is documented in docs/CLI.md; the recommended commands to run after each kind of edit are in docs/AGENT_COMMANDS.md. See VALIDATION.md for the full list of checks.

Platform pack manifests and plugin-compatibility guidance are documented in docs/PLUGIN_COMPATIBILITY.md. The generated metadata files metadata/index.json, metadata/router.json, and metadata/packs.json are the machine-readable surfaces for skills, routing, and platform packs.

Contributing

New skills and improvements are welcome. AgentCounsel is Markdown-first and safety-first — see CONTRIBUTING.md for the rules (including how to add skills, quality checks, evals, playbooks, review panels, and packs), docs/SKILL_METADATA_STANDARD.md for the frontmatter standard, and SECURITY.md for security guidance. Questions are answered in docs/FAQ.md.

For the bigger picture, see docs/PROJECT_STATUS.md (an honest maturity snapshot — what is stable, what needs manual review, what is experimental) and docs/WORKFLOW_MAP.md (how core rules, skills, quality checks, packs, workspaces, playbooks, review panels, and evals fit together).

License

MIT License — see LICENSE. Attribution for content adapted from other open-source projects (which remain under their original Apache-2.0 terms) is recorded in THIRD_PARTY_NOTICES.md. AgentCounsel is an independent open-source project and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, or Google.

Disclaimer

AgentCounsel is a workflow library — not a law firm and not a lawyer. Using it does not create an attorney–client relationship, and it does not provide legal advice. All output is draft work product that must be reviewed and adopted by a qualified, licensed legal professional before it is relied upon. You are responsible for the use of these materials and for compliance with the rules of professional conduct that apply to you.