correct
Use when the human wants a mistake agents keep making fixed. Starts from their correction or the mistake they name, finds what led the agent there, and fixes it where the cause is: the project's docs, the code, or the prompt's source; a cause in an installed prompt becomes a note for its owner.
How do I install this agent skill?
npx skills add https://github.com/haowjy/creative-writing-skills --skill correctIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill modifies project instructions and documentation based on historical session data to fix recurring agent errors. This creates a surface for indirect prompt injection where past interactions could influence the agent to persistently alter its own rules or security configurations.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Correct
The human corrected an agent and doesn't want to do it again. Find what led the agent to the mistake, and change that.
Find the Cause
Start from the correction the human gave or the mistake they name. Read the
session where it happened: what the agent read, what it was told, what it
did (@session-miner finds the passages; read them yourself). When the
human asks for a sweep, look for the same mistake in recent sessions,
reverts and review comments.
A single occurrence can be a bad run, with nothing in the setup to blame. When that's what the evidence shows, say so and change nothing.
The usual causes:
- a line the agent read that misled it: stale, contradicted by the code, or buried under lines that matter less
- code that invites the mistake: two ways to do one task, a misleading name, a dead example that copies wrong
- a prompt (an agent or a skill) that steers wrong, or says nothing at the moment of choice
The same cause usually sits in more than one place: a stale rule copied
into other AGENTS.md files or KB pages, the same pattern in sibling code,
the same wording in another prompt. Search for it; the cause is found when
every copy is.
Fix It Where the Cause Is
- A misleading line in
AGENTS.md,.context/or the KB: correct or delete it. Thinning is often the fix: a rule lost among ten others gets followed once they go. Add a line only when nothing the agent read covers the choice, and put it where the agent reads it as it makes that choice. - Code that invites the mistake: make the wrong way stop working, strongest first. One owner per piece of state and one supported way per task, with the old way and its dead examples deleted; then types, so the bad state can't be written; then a lint whose error names what to use instead, proven to fail on a real past mistake. This is a code change: it goes through the project's normal flow, not in with the correction.
- A prompt: fix it at its source. When the source is the repository
you're working in, edit it there for the owner to review. When it's an
installed package, leave the installed copy alone, since the next sync
overwrites it: leave a note where the package's owners will see it
(wherever its
AGENTS.mdsays, an issue on its repository, or a message to whoever maintains it) saying what the human corrected, in which session, and the line you suspect.
A permission or hook that blocks an agent's action is for a destructive mistake that has repeated, with the human's approval: it blocks legitimate work too.
Report
Each mistake with its evidence and cause, and what you changed or where you left the note; and each one you judged a bad run, with why.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/haowjy/creative-writing-skills/correct">View correct on skillZs</a>