log-correction

Skill: Log Correction

Purpose

Record analyst mistakes, their fixes, and reusable learnings so future analyses learn from past errors. Runs in two modes against the same store (defined in docs/KNOWLEDGE.md): auto mode detects corrections and learnings in the user's messages without being asked, and manual mode handles explicit "log a correction" requests with full detail.

When to Use

• Auto: the user corrects your work ("that's wrong", "actually it's...", "you used the wrong column") or teaches a reusable rule ("always use X", "never do Y", "remember that our fiscal year starts in February") without asking you to log anything
• Manual: user says "log a correction", "save this mistake", "record this lesson", or similar
• After discovering and fixing an error mid-analysis worth preserving

Auto Mode

Watch every user message for these signals. When one fires, capture it immediately; the user never has to ask.

Correction signals (something you produced was wrong):

• "that's wrong", "that's incorrect", "actually it's...", "it should be..."
• "the column is X not Y", "you used the wrong...", "off by...", "double-counted", "that join is wrong", "missing a filter", "forgot to exclude..."

Learning signals (a reusable methodology or fact):

• "always use...", "never use...", "next time...", "prefer X over Y"
• "remember that...", "the convention here is...", "our team uses...", "going forward...", "don't forget to..."

If both match, treat it as a correction. If neither matches, do nothing and say nothing about it.

On a correction signal: run Steps 1-5 below, but never interrogate the user. Infer severity, category, dataset, and tables from context; leave fields you cannot infer as null. Acknowledge in one line ("Got it, logged as CORR-008.") and then immediately continue with the user's underlying request; logging is never the whole response.

On a learning signal: append a bullet to .knowledge/learnings/index.md under the closest category heading (Data Patterns, Query Techniques, Business Context, Stakeholder Preferences, Visualization Insights, Methodology Notes), formatted - {concise learning} (source: user feedback, {YYYY-MM-DD}). Acknowledge in one line ("Noted for future analyses.") and continue with the user's request.