The damage starts before Claude writes code.
AIDE catches your weak prompts in real time and fixes them before the agent acts.
<request> Do not repeat the previous approach. Diagnose why the last attempt failed, then apply the smallest fix and verify. </request>
<original_prompt> fix the auth middleware </original_prompt>
Learn offline. Coach live.
Install now, coached immediately. Mining your history is optional and makes the coaching yours.
One command
npx agentaide registers the hooks. No dependencies, no config, no account.
Coached from the first prompt
21 built-in rules catch blind retries, error dumps, stale sessions, and runaway tool spirals. Deterministic, in milliseconds.
Gets personal over time
Optionally mine your own history into a rulebook.json: your patterns, not generic advice, re-tuned automatically as you work.
The mistakes you don't notice until the bill arrives.
Tokens, quality, and risk. The hidden cost structure of agent-assisted development.
Stale session resumption
Starting a new task in a 2-day-old chat silently re-reads 30k tokens of dead context every single turn.
Blind retries
Re-sending the same prompt that just failed tends to produce the same failure, plus another correction loop.
Web-fetch spirals
The agent fetches page after page instead of using what it already has. AIDE steers first; full scope also denies the tool mid-spiral.
Shipping without tests
“Commit and push” with zero test evidence this session gets a verification packet injected before the agent acts.
Error dumps without context
A pasted stack trace with no repro steps sends the agent guessing. AIDE wraps it so the agent diagnoses before editing.
Destructive commands on autopilot
With full scope on, a destructive bash command under auto-accept gets caught before it runs, not after.
Four channels. One fatigue budget.
Max one visible notice per prompt, three per session. A coach that nags gets muted. AIDE is built not to.
Silent steer the agent reads alongside your prompt. You see nothing; the agent course-corrects.
A one-line visible nudge with evidence: token counts, idle hours, correction history.
Your prompt passes through; the agent acts on an optimized rewrite. You see one line: “prompt optimised”.
Opt-in full scope. Denies a runaway tool call mid-spiral: web chains, bash retry loops, destructive commands.
Local by design. Not surveillance.
Your transcripts, your rulebook, your telemetry: everything lives in ~/.claude-judge/ on your disk. Nothing phones home, and no manager reads your prompts. Optional prompt polish runs through your own Claude CLI or API key.
Your team's Claude Code bill has a leaderboard. So does its skill.
Some engineers get 10x more out of coding agents than others sitting next to them. The difference is learnable, if you can see it.
- See how your best engineers work. How they decompose tasks, when they clear context, how they recover from failures. Measured from real sessions, not self-reports.
- Spread the habits, not the blame. Everyone gets the personal coach. The patterns that make your top performers effective become the nudges everyone else receives.
- Cut the waste you can't see today. Token spend, correction loops, and unverified ships, quantified per pattern and improved per person.