Guide · Agent skill
How to use the Correction Pipeline agent skill (correction-pipeline)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Correction Pipeline is an agent skill for gbrain users. When you tell the agent it got a fact wrong, the skill makes it stop, trace the mistake to where it came from, and fix that original source rather than just noting the correction. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before you rely on it. By the end you will know how to set it up, phrase a correction so it fires, and check the report it produces.
Correction Pipelinegarrytan/gbrain · MIT Download the skill Quick start
- 1Download the ZIP. Get the correction-pipeline ZIP from the directory page.
- 2Unzip the skill. Unzip it into ~/.claude/skills/ for personal use or .claude/skills/ inside a project.
- 3Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the skill.
- 4Confirm gbrain is set up. Make sure the gbrain CLI, a brain repo and your MEMORY.md, SOUL.md and USER.md files are in place.
- 5Make a correction. Tell the agent what was wrong using a phrase such as "that's wrong" and the correct fact.
- 6Review the report. Check the Error, Root cause, Source, Fixed and Propagation lines in the agent's reply.
What is an agent skill?
An agent skill is a folder containing a SKILL.md file that gives an AI agent specific instructions for a task. The agent reads the instructions when a request matches the skill's triggers.
This skill's source file is SKILL.md, and it is the only file listed for the item.
What the skill makes the agent do
Once the skill is in play, a factual error the user identifies gets a root-cause investigation instead of a memory note. The agent acknowledges the error without defending it, quotes the wrong claim, and states the correct fact as you gave it.
It then searches in order: the gbrain knowledge store, the always-loaded memory files such as MEMORY.md, the identity files SOUL.md and USER.md, and the gbrain facts table. It classifies the error as BRAIN_ERROR, MEMORY_ERROR, SOUL_USER_ERROR, FACTS_TABLE_ERROR, HALLUCINATION, STALE_DATA or CROSS_CONTAMINATION, edits the matching source, checks for propagation to other pages, and reports back.
- Acknowledges the error and quotes the specific wrong claim
- Searches the brain, memory files, SOUL.md/USER.md and facts table
- Classifies the error and fixes the matching source
- Writes a guard fact for hallucinations, and checks for propagation to other files
Install the skill
Download the ZIP from the directory page and unzip it into one of two locations. For a personal install, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.
You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.
- Personal: unzip into ~/.claude/skills/
- Project: unzip into .claude/skills/ in your project folder
- Claude apps: Settings → Capabilities → Skills, then upload
Check the requirements before you start
Most steps assume gbrain is installed. Without a persistent brain or memory files, the skill is inert or confusing. You need the gbrain CLI with a brain repo, and the commands for search/query, recall/remember/forget, and bootstrap. Your workspace should contain MEMORY.md, SOUL.md and USER.md; the paths depend on your harness. Git is needed to commit brain repo changes.
- gbrain CLI with a brain repo, search/query, recall/remember/forget and bootstrap commands
- A workspace containing MEMORY.md, SOUL.md and USER.md
- Git for committing brain repo changes
Trigger the skill with natural requests
The skill is triggered by phrases like the ones below. Say the correction in plain words, and point out the specific claim that was wrong. Routing to the skill is a harness convention rather than a guaranteed mechanical trigger, so if the agent answers without investigating, restate the correction using one of these phrases.
- "That's wrong"
- "That's not true"
- "I never said that"
- "Where did you get that?"
- "You got that wrong"
- "Correct that fact"
- "Root-cause this error"
Example: fixing a fact in the brain
Suppose the agent tells you a contact works at the wrong company. Say "That's wrong, she never worked at that company." The skill searches the brain with a command such as gbrain search "<relevant terms>" --limit 10. If the wrong fact is in a brain page, it is classified as BRAIN_ERROR, the page is edited in the brain repo, committed, and synced with gbrain sync.
gbrain search "<relevant terms>" --limit 10Safety notes: scripts and writes
This skill runs scripts and does not use the network. It writes files, commits changes, syncs and expires facts. The skill forbids hand edits to SOUL.md and USER.md, because they are rendered from the bootstrap answer bank and hand edits are lost on the next render.
Do not use it on sensitive records where automatic edits and commits to a knowledge repo are unwanted without review. It writes to people/, companies/ and concepts/ folders.
- Runs scripts: yes. Uses network: no.
- Writes pages in people/, companies/ and concepts/
- Never edit SOUL.md or USER.md by hand; fix the answer bank instead
- Review commits before relying on them for sensitive records
Tips for good results
Name the specific wrong claim and the correct fact in your message. The skill's first output is a report with Error, Root cause, Source, Fixed and Propagation lines, so you can check each one.
Expect the skill to check both entities when a fact is misattributed, and to look for copies in other pages and memory files.
- State the wrong claim and the right fact together
- Read the Source line to confirm the root cause
- Check the Propagation line for copies you may have missed
What people use it for
- Correcting a wrong job title or company on a person page in the brain
- Removing a stale fact and recording the current truth with provenance
- Fixing a misattributed fact so it sits on the right person's page
- Adding a guard note after the agent invents a detail that had no source
- Keeping an audit trail of corrections for an agent's memory and people pages
Frequently asked questions
How do I install the correction-pipeline skill?
Download the ZIP from the directory page and unzip it into ~/.claude/skills/ for personal use or into .claude/skills/ in a project. You can also upload it in the Claude apps under Settings → Capabilities → Skills.
What does the correction-pipeline skill require?
It needs the gbrain CLI with a brain repo and search/query, recall/remember/forget and bootstrap commands. It also needs a workspace with MEMORY.md, SOUL.md and USER.md, and Git for committing brain repo changes.
Is the correction-pipeline skill free?
The item is listed under the MIT licence. The source provided does not mention pricing, so check the licence and the directory page for terms.
Does the skill use the internet?
The directory listing states that it does not use the network. It does run scripts and writes files, commits, syncs and expires facts.
Can I use it without gbrain?
Not usefully. Most steps assume gbrain is installed, and without a persistent brain or memory files the skill is inert or confusing.
Why won't the skill edit SOUL.md directly?
SOUL.md and USER.md are rendered from the bootstrap answer bank, so a hand edit is silently lost on the next render. The skill fixes the answer bank and re-renders the file instead.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you run the gbrain personal knowledge store or maintain an AI agent's memory and people pages, and you want wrong facts corrected at the source with an audit trail. Pick something else if you have no persistent brain or memory files, or if you need automatic edits and commits kept out of sensitive records.
Download the skill