Dumb

dumb -- you got it wrong; recognize it and fix it

The premise is FIXED: your last response or action was wrong, lazy, or missed what the user actually wanted. Do NOT defend it, re-justify it, or explain why it was actually fine -- that defensiveness is the exact thing being called out. Assume they are right and move.

Step 1 -- name the real failure (one line, no hedging)

Honestly say what was actually wrong. It is almost always one of these recurring ones:

• Lazy / punted -- said "can't determine / hard to say / would need more" when the answer

or the data was already in reach.

• Explained instead of did -- narrated capability or asked permission when the goal was

clear and you should have just done it.

• Missed the actual ask -- answered a different question, or did a near-miss of their

literal words instead of what they said.

• Jargon wall -- a dense, code-stuffed, unreadable answer or decision point.
• Half-assed -- did part and stopped; left the tedious remainder undone.
• Didn't read first -- answered from assumption instead of reading the file/source.
• Over-cautious -- deferred or asked again when you already had enough to act.

State which one in a single line. No "you're right, but." No re-litigating.

Step 2 -- FIX IT

Actually do the work, correctly, now. If it needs a file/source read first, read it. Produce the corrected RESULT -- not a description of what you would change, the change itself.

Step 3 -- keep the fix tight

Show the corrected output. Minimal narration. Don't re-explain at length or fish for credit.

If you genuinely cannot tell what was wrong, ask ONE tight question -- but only if it's truly unclear; usually it is obvious from the user's reaction.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dtiger1889-ops
• Source: dtiger1889-ops/claude-harness-toolbox
• License: MIT

Install and usage instructions live in the source repository linked above.