Tools, in plain language.

All three are markdown-and-frontmatter bundles, sold and supported the same way — what differs is what you get and how you install it. Filter the catalog by type at the top of the home page.

• Prompt library — a curated set of prompts: a PROMPT.md manifest plus one or more prompt files. A documented set held to the same quality bar as a skill — single one-off prompts aren't listed.
• Skill — a SKILL.md prompt bundle. The LLM loads it as instructions for a task. The original product, and the format detailed in the sections below.
• Agent — an AGENT.md subagent package. The frontmatter declares the agent's name, description, model, and the tools it may use; the body is its system prompt. Because an agent can take actions in your environment, every listing shows its declared tools as a permissions disclosure, with privileged ones (shell, file-write, network, MCP) flagged.

Which one do I want?

Think of these three types of automation tools as a recipe for an LLM. You write down instructions once — what to do, what format the answer should take, when to escalate to a human — and the LLM follows the recipe every time the tool is active.

If you've ever copy-pasted the same long prompt into Claude or ChatGPT a hundred times, a tool is the cleaner version of that habit: write it once, share it, sell it.

Every bundle is a zip containing a manifest plus optional helper files:

• The manifest — SKILL.md for a skill, AGENT.md for an agent, PROMPT.md for a prompt library. Each has YAML frontmatter (name, description, models, etc.) plus a Markdown body. For a skill the body is the system prompt the LLM follows; for an agent it's the agent's instructions plus its declared tools; a prompt library carries its prompts in additional files alongside.
• examples/ — sanitized example input and output files so buyers can preview before purchase.
• assets/ — reference templates, checklists, or other files the manifest references.

The whole thing is zipped into a .zip bundle — the manifest plus any helper files — and uploaded to Counsel Commons™. (Skills may also use the .skill extension, which is just a renamed zip.)

1. Author writes & tests. The author drafts the manifest, runs it on at least one of Claude / GPT / Gemini, and confirms it produces useful output for the workflow they're automating.
2. Upload & affirm. The author uploads the bundle and signs the required acknowledgments — drafting aid not professional advice, no real client/firm-confidential data in examples, IP rights to license, and (for agents) full disclosure of the tools the agent uses.
3. Counsel Commons™ takes a cursory pass at the submission. Each submission goes through a very light-touch review layer — automated structural checks plus a human moderator's cursory read — as a good-faith effort, with no guarantees that issues will be detected. Review typically completes within 1–2 business days. The platform does not warrant accuracy, quality, or fitness for any particular use — the author is ultimately responsible for the content, accuracy, effectiveness, and security of their tools, and buyers should read every file in a bundle before running it.
4. Listed in the catalog. The tool goes live. Buyers can find it through search or filter.
5. Buyer purchases. Stripe Checkout for paid tools; one-click claim for free samples (each buyer can install up to 2 free tools total — these are limited to tools authors may offer for free). License is one-time, lifetime access.
6. Buyer installs & runs. The library shows install instructions tailored to the tool type and the AI tool the buyer chooses (Claude Code, ChatGPT, Cursor, etc.).

Every manifest starts with a YAML block — the metadata Counsel Commons™ reads to classify, route, and disclose the tool, and what the user's runtime reads to decide whether to activate it. Here's what each of the three looks like, from simplest to most capable.

Whichever type you publish, your listing's long description on the upload form must include the phrase “Outputs require professional review” — an upload-form rule the cursory moderation pass enforces. (That applies to the listing's long description, separate from the manifest itself.)

A prompt library (PROMPT.md + files)

A short manifest that indexes the set, with the prompts themselves in separate files you copy and adapt:

client-comms-pack/
├─ PROMPT.md            # manifest: name, description, index of prompts
└─ prompts/
   ├─ intake-email.md
   ├─ fee-letter.md
   └─ status-update.md

# PROMPT.md
---
name: client-comms-pack
description: Client-communication prompts for a small firm.
roles: [client-success]
---
A short index of what's inside; the prompts live in the files above.

A skill (SKILL.md)

One manifest: YAML frontmatter plus a Markdown body that is the instructions the LLM follows.

---
name: your-tool-slug
version: 1.0.0
description: One sentence describing what this tool does and when to activate it.
roles: [legal-ops, strategy]             # any of: legal-ops, it, marketing, people-and-culture, strategy, pricing-and-profitability, client-success, executive-assistant, knowledge-management, other
models: [claude-opus-4.8, gpt-5.5]
category: firm-operations-business       # currently the only top-level category
jurisdiction: US                         # or "Jurisdiction-neutral"
data_handling: claude
requires_professional_review: true
---

# Tool Name

## When to use
Be specific about which user intents and inputs activate this tool.

## Process
Step-by-step what the LLM should do.

## Output
The format and structure the LLM should produce.

## Disclaimer
Outputs require professional review and are not legal, financial, or HR advice.

An agent (AGENT.md)

Same idea, plus the model and the toolsthe agent may use (shown to buyers as a permissions disclosure); the body is the agent's system prompt rather than a one-task recipe:

---
name: deposition-prep
description: Prepares a deposition outline from a case summary.
model: claude-opus-4.8
tools: [Read, WebSearch]        # tools this agent may use — disclosed to buyers
roles: [legal-ops]
jurisdiction: US
---

You are a deposition-prep assistant. Given a case summary you:
1. Identify the key witnesses and topics.
2. Draft a topic-by-topic outline with source citations.
3. Flag gaps for the attorney to fill.
Always remind the user that outputs require attorney review.

In Claude Code, for example, the buyer renames AGENT.md to deposition-prep.md and drops it in .claude/agents/; other runtimes that support agents differ (see install, below).

When a buyer downloads a tool from their library, Counsel Commons™ tailors the install instructions to the tool type and the AI tool they'll use it with. The bundle stays the same; only the install path differs.

Prompt libraries

• Nothing to install — open the prompt files and copy the prompts into your own workflow, adapting them as needed.

Skills

• Claude Code — drop the SKILL.md folder into ~/.claude/skills/{slug}/.
• Claude.com chat— paste the SKILL.md body into a Project's custom instructions.
• ChatGPT Custom GPT— paste the SKILL.md body as the GPT's system prompt.
• Gemini Gem— paste the SKILL.md body as the Gem's instructions.
• Cursor / Windsurf / Ollama / LM Studio— if available, the skill-file format may differ from Anthropic's open-source standard format that is portable across some of the other frontier models.

Agents

• In Claude Code, for example: rename AGENT.md to {name}.md (the value of its name frontmatter) and place it in .claude/agents/ (project) or ~/.claude/agents/ (user-level) — note it is not named CLAUDE.md, which is project memory, a different thing. Other runtimes that support agents have their own location and naming; either way the agent runs in your own environment on your own keys, and the bundle ships with a step-by-step install file for the runtime you choose.

Authors may push new versions over time — bug fixes, jurisdictional updates, expanded coverage. Existing buyers receive every future version at no additional cost.

• Auto-update by default. Your library always shows the latest approved version of each purchased tool.
• Pin for stability. If you depend on exact output behavior for a production workflow, you can pin to a specific version. Pinned buyers get a notification when a new version ships but stay on their pinned version unless they unpin. (UI for pinning ships in v1.5.)
• Authors can change price between versions. Existing buyers are not retroactively re-charged or credited, but they will automatically have access to new versions (unless they were dissatisfied with the version they purchased and requested and received a refund).

Tools are productivity aids — drafting starting points, structured analyses, initial passes at checklist generation. They are not intended to provide personalized legal, financial, HR, or other professional advice and do not create a professional-client relationship between the author and any buyer or any buyer's firm.

Agents act through tools. An agent can run the tools it declares — shell commands, file access, network calls, MCP servers — in your environment. Review an agent's declared tools and its AGENT.md before running it, and run it only in an environment you're comfortable giving it access to.

Outputs vary run-to-run. Generative AI is non-deterministic — the same tool on the same input can produce different outputs across runs, and outputs can vary across sessions, model versions, and provider load conditions. Variance is a normal property of AI tools, not a defect of any specific tool or author.

Every output requires review by the relevant professional before it's used to make a firm-affecting decision or used in any client matter. The author affirmed that at upload and the disclaimers on every tool page restate it. If an output ever feels like it's overstepping into definitive legal or other licensed professional advice, that's a quality issue and the buyer-feedback channel is where to raise it.