How to use the Proof markdown editor agent skill (ce-proof)

What the ce-proof skill makes the agent do

An agent skill is a folder of instructions that an AI agent loads when a request matches the skill's description. The ce-proof skill teaches the agent how to talk to the Proof web API over HTTP.

With the skill, the agent can publish a local markdown file to Proof and give you a shareable link. It can also read a shared Proof document, add comments or suggested edits, reply to or resolve review comments, and delete documents when you ask.

Publishing is one-way. The local file stays the source of truth, and nothing is synced back to disk.

• Publish a local markdown file as a new Proof document and return the link
• Read a shared Proof document
• Add comments or suggested edits, reply to comments, and resolve them
• Delete a document when you ask for it

Install the ce-proof skill

Download the ZIP of the ce-proof skill from this page and unzip it. The skill folder should contain SKILL.md and a references folder with api.md and workflows.md.

For a personal install, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder so it applies only there.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Per project: .claude/skills/ inside the project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Check what the skill needs

The skill needs network access to https://www.proofeditor.ai. It uses the Bash tool to make HTTP calls with curl, and it uses jq in the publish recipe. If your harness provides typed Proof MCP tools such as proof_share_markdown or proof_v3_edit, the agent prefers those instead of raw HTTP.

You do not need a separate account for anonymous publishing. Access tokens and owner secrets are created for each document at publish time.

• Network access to https://www.proofeditor.ai
• Bash tool for curl, and jq for the publish recipe
• Optional: Proof MCP tools if your harness provides them

Trigger the skill with natural requests

You do not need a special command. Ask in plain language and name the file or link. The skill's description covers Proof links, sharing specs, plans and drafts, and publish handoffs from planning workflows. The agent asks which file you mean only if the request is ambiguous.

The skill also receives handoffs from the ce-plan, ce-brainstorm and ce-ideate skills, which pass along a file path and a title.

• "Share this to Proof" with a markdown file open or named
• "Get me a Proof link for this spec"
• "Read this Proof link and tell me what the reviewers commented on"
• "Add a comment on the rollout section of this Proof doc asking about the timeline"

Understand the publish and review flows

When you publish, the agent reads the whole local markdown file and posts its exact contents as the new document body. It then gives you the tokenized link. Hand that link to reviewers, not the bare document address, because the token is what grants access.

When you review a shared document, the agent reads the current document state first, then makes narrow edits. Comments and suggestions are handled as review operations, so reviewers see the changes in context. The agent reports results only after it has confirmed them by re-reading the document where needed.

Safety notes before you share

Anyone with the share link can read the document. Do not publish secrets, credentials, API keys, private tokens or sensitive personal data unless you explicitly approve it.

Clearing the text of a document does not remove comment text. Deleting the document is the only real cleanup. While a document is unclaimed, deletion needs the owner secret the agent stored at creation. After a signed-in Every user claims the document, deletion needs that owner's session.

The skill does not auto-delete after a publish. Review documents stay up until you ask for deletion.

The skill runs no scripts. It makes network calls to proofeditor.ai, and pulling a document to a local file overwrites that file, so the agent confirms the path first when the pull was not something you asked for.

• Share links grant read access to anyone who has them
• Emptying a doc is not a privacy cleanup; delete it instead
• Keep the local markdown file as your master copy

Tips for better results

Give the agent a specific file path or a clear description of the document. Ask for a title if you want a particular one; for unified plans the agent labels titles by readiness when that is known.

If you want quieter sessions on Claude Code, the api.md reference suggests an allowlist rule for curl calls to www.proofeditor.ai. The skill says not to add it silently, so only add it if you choose to.

If a publish or edit fails, ask the agent to re-read the document before retrying. Ask it to report a bug rather than loop when a fresh read and one safe retry still fail.

• Name the exact file to avoid a clarifying question
• Re-read the document before retrying a partial write
• Use the optional allowlist rule only if you want fewer permission prompts