How to use the Skillpack Harvest agent skill

What the skill makes the agent do

Harvest lifts a proven skill from a host repository into gbrain's bundle. The gbrain CLI does the file copy and the privacy lint. The agent handles the judgment-heavy part: replacing real names, people, companies and internal channels with generic placeholders, generalizing trigger phrases, and moving fork-specific conventions into references.

The skill is the reverse of gbrain's scaffold command. Scaffold ships skills downstream from gbrain to a host. Harvest sends proven patterns upstream from a host to gbrain, so other clients can scaffold them.

• Copies the skill into gbrain's working tree under skills/<slug>/
• Can include paired source files listed in the host SKILL.md frontmatter sources: array
• Adds the new slug to skills: in openclaw.plugin.json (sorted)
• Runs a dry-run preview and a privacy lint before the real harvest
• Leaves the result uncommitted until you review the diff and approve the commit

Install the skill in Claude

Download the skillpack-harvest ZIP from the directory page. Unzip it into your user skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project.

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

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into .claude/skills/ inside the project
• App install: upload the ZIP under Settings → Capabilities → Skills

Trigger the skill with natural requests

You do not need a special command. Ask in plain language, and the agent will match your request against the skill's trigger phrases. Name the skill you want to harvest, and the agent will ask for the slug and the host repository path if it does not already have them.

The source lists these trigger phrases: "harvest this skill", "harvest my skill", "publish this skill to gbrain", "lift this skill", "share this skill", "promote this skill", "promote my skill", "skill upstream", "into the gbrain core" and "gbrain bundle".

• "Harvest my skill called invoice-triage from ~/git/agent-fork into gbrain."
• "Promote this skill to gbrain. Start with the dry run."
• "Share this skill upstream, but scrub my names and company first."

Run a harvest, step by step

Phase 1 is planning. The agent asks for a kebab-case slug that is unique in the gbrain bundle, the path to the host repository root (not the skill folder), and whether paired source files should come along.

Phase 2 is the dry run. It shows which files would land in gbrain's tree and whether paired sources are included. Phase 3 is the editorial pass, where the agent walks the host's skill files and applies a checklist. Phase 4 is the real harvest, which runs only after Phase 3 is complete.

• Phase 3 covers names, real entities, email addresses, Slack channels and ticket IDs
• It also checks that triggers do not mention your name, fork name or internal tools
• routing-eval.jsonl intent fields get the same scrub as triggers
• Comments and log strings in paired source files are checked for private details

Safety notes: scripts and network

This skill is marked as mutating and runs scripts, so it changes files in your gbrain checkout. The source says harvest is a copy, so the host keeps its skill. It does not declare network use. Harvest writes into gbrain's working tree, and nothing counts as landed until the user has reviewed the diff and approved the commit.

If the harvest finishes with status harvested and the privacy lint reports no hits, the skill still is not done until the conformance test passes and the user has approved the commit.

• Files land in the working tree, so a git checkout can undo them if you change your mind
• Do not run the harvest into a gbrain checkout you do not intend to commit from
• Do not remove the source skill after harvesting
• Harvesting with --no-lint is discouraged; if you must, document the reason in the commit

Tips and common mistakes

Harvest a skill only after it has proven itself. The source asks that recent routing-eval cases pass and that the skill has been used in production a few times. Harvesting a skill that is still changing causes repeated re-harvests and churn in gbrain's bundle.

If the linter catches a real name or private detail, go back to Phase 3 and replace it with placeholders such as alice-example, acme-example or fund-a. Do not rely on the linter alone, because its default regex set does not catch every proper noun.

• Run the dry run every time, even for small edits
• Strip-test the skill in your head: would it still make sense with every fork-specific name replaced?
• If skillpack check --strict reports drift after harvesting, regenerate the bundle and re-run rather than shipping a drifting skillpack
• Batch harvesting several skills at once is not supported

Test the harvested skill

After the real harvest, the source requires that the conformance test pass on the new SKILL.md. Bun is needed to run the test file test/skills-conformance.test.ts inside the gbrain checkout. If the test fails, fix the file and run the check again before asking for commit approval.