How I Stopped My LLM from Wasting Request Quota with Annoying Confirmations

I was pulling my hair out. My LLM in agent mode kept pausing to ask “Should I proceed?” before every tiny action, burning through my request quota like a kid with a credit card. After weeks of tweaking and swearing at my screen, I finally cracked the code to make it stop wasting tokens on useless confirmations.

The Core Problem

Every time I asked my AI to fix code or update files, it responded with tutorials or confirmation questions instead of just doing the job. Each “Are you sure?” wasted a precious request. I’d have to reply “yes” just to keep things moving. Felt like paying highway tolls just to check my mailbox.

My Initial Struggles

I tried adding strict rules like “Never ask for confirmation!” but the behavior persisted. Testing different models in Auto mode (GPT-4.1, Deepseek) backfired – they’d default to chatty lectures. At one point I even wondered if this was some quota-grabbing scheme! Turns out the real issue was how I talked to the AI.

What Actually Worked

• Model choice matters: Claude models (especially Claude 4) act directly when given tasks, unlike GPT-4.1’s textbook responses
• Precise prompts win: Vague requests trigger explanations; surgical commands get immediate action
• Stay positive: Saying “Don’t waste requests” confuses the AI – positive phrasing like “Execute immediately” works better
• Trim context fat: Overloading with files or rules distracts the model – keep supporting docs lean and task-focused

My Step-by-Step Fix

• Switched from Auto mode to Claude-only for agent tasks – its “doer” mentality was key
• Radically simplified prompts:
  • Changed “What’s wrong here?” → “Fix the schema in dashboard.py”
  • Replaced “Analyze this bug” → “Patch the login error in auth_service.js”
• Optimized context files:
  • Kept memory.md razor-focused on progress tracking
  • Used clear action commands like:
    Implement requirements directly without confirmation
Test changes using Playwright MCP automatically
• Tested every new prompt in sandbox mode first to verify immediate execution

The Results and What I’d Tell You

After these changes? My confirmation loops dropped 80% overnight. Now when I say “Update the API endpoint,” it just does it – no more token-wasting chatter. If you’re fighting this battle, remember: be surgical with your wording and match your model to the job. Claude + direct commands saved my sanity and quota.