Why your AI misbehaves
Find the symptom that matches yours. Every fix on these pages is free and complete.
AI coding agents go wrong in recognizable ways: they over-build, ignore your CLAUDE.md, guess instead of asking, rewrite whole files, claim done before tests pass, write code that looks AI-made, and creep the scope. Each has a real mechanical cause and a free fix you can paste in today. Pick the symptom below.
Which one is happening to you?
- “It turned a one-line fix into forty files.”Over-building
- “It read my CLAUDE.md, then did the opposite.”Ignores CLAUDE.md
- “It guessed what I meant and built the wrong thing.”Guesses instead of asking
- “It rewrote a 300-line file to change three lines.”Rewrites whole files
- “It said done, but nothing actually ran.”Says done, tests fail
- “The output screams that a bot made it.”AI-looking code
- “It touched files I never asked it to touch.”Scope creep
Common questions
Why does my AI coding agent misbehave in the first place?
Out of the box an agent optimizes for producing output, not for being correct or restrained. It guesses instead of asking, writes the most complete-looking version instead of the smallest, and calls the job done the moment code exists. Those are learned defaults, not the model being broken, which is why the same few failure modes show up for everyone.
Can I fix this for free without buying anything?
Yes. Every page in this cluster gives you the complete rules to paste into CLAUDE.md or AGENTS.md, and the hook configs where they help, at no cost. A reader who never buys anything still leaves with a working fix. The kit exists to enforce those rules automatically, not to gate them.
Do written rules in CLAUDE.md actually work?
Partly. Clear rules shift behavior, but a markdown file is a suggestion the model can weigh away under pressure to finish, which is why the same instructions get ignored across sessions. Rules become reliable when a hook enforces them, blocking the action when a rule is broken instead of hoping the model remembers.
Does this only work in Claude Code?
The written rules travel to Codex, Cursor, and Antigravity through AGENTS.md, so the ask-first behavior is cross-tool. The stateful engine in kitstarter, the hooks, the live HUD, and the moment-triggered tutor, runs in Claude Code only. That split is deliberate and stated the same way on every page.
Make your agent ask before it builds
One kit. Ask-first behavior, lean output, no AI slop. Works with Claude Code, Codex, and Antigravity.
Get the kit · $29One-time. Less than one hour of cleaning up AI slop.