UNPKG

major-ai-skills

Version:

Installable agentic skills / AI agent skills (SKILL.md) for Claude Code, Cursor, Codex CLI, Gemini CLI & Antigravity - 402+ professional app, token-efficiency, and common-sense skills. SEO/GEO ready.

91 lines (71 loc) 4.78 kB
--- name: stop-wrong-answers-early description: "Interrupt a clearly misdirected answer and provide a specific correction to scope or assumptions." category: common-sense risk: safe source: self source_type: self date_added: "2026-08-26" tags: ["fast-abort", "token-savings", "context-hygiene", "stop-generating", "productivity", "prompt-engineering"] tools: ["claude", "cursor", "gemini", "codex", "chatgpt"] --- # Stop Wrong Answers Early (Fast Abort & Edit) (AI Skill) ## Overview When users see an AI start generating in the wrong direction (*e.g., generating code in Java when they wanted Python, or writing a 500-word essay when they wanted 3 bullets*), many passively wait 30 seconds for the model to finish before replying: *"No, that's wrong."* Passively waiting wastes time, burns output token credits, and **pollutes the conversation history** with a giant block of incorrect text that will skew subsequent turns. The **Fast Abort & Edit Protocol** makes hitting the **"Stop Generating"** button an instinctive, real-time action, immediately editing the source prompt to correct the trajectory. --- ## Passive Waiting vs. Fast Abort & Edit ``` ┌─────────────────────────────────────────────────────────────┐ Passive Waiting vs. Fast Abort Passive Waiting: Watch model write 600 tokens in the wrong direction Wait 30 seconds Type new message: "No, in Python not Java" Result: 1,500 total tokens billed, polluted context Fast Abort & Edit (The 3-Second Rule): See Java syntax $\rightarrow$ HIT STOP in 2 seconds Click "Edit Prompt" $\rightarrow$ Append "in Python 3.12" Re-run $\rightarrow$ Perfect Python output instantly Result: Zero wasted tokens, clean pristine context └─────────────────────────────────────────────────────────────┘ ``` --- ## The 3-Step Abort & Edit Reflex ``` ┌───────────────────────────────────────────────────────────────────────────┐ 1. DETECT (Sec 0-3) ──► Spot incorrect language, wrong format, or fluff 2. ABORT (Sec 3) ──► Hit [STOP GENERATING] / [CANCEL] immediately 3. EDIT SOURCE ──► Click the pencil icon on YOUR original prompt, add the missing constraint, and re-submit └───────────────────────────────────────────────────────────────────────────┘ ``` --- ## Why Editing the Original Prompt Beats Sending a Correction When you hit **Edit** on your original prompt instead of sending a new message: 1. **Erases the Mistake**: The incorrect output is permanently removed from the conversation tree rather than remaining in context history. 2. **Saves Input Tokens**: Prevents the failed generation from being re-billed on every future turn. 3. **Eliminates Model Confusion**: Prevents the model from trying to reconcile contradictory messages in the same thread. --- ## Real-World Case Study ### Scenario: Requesting a Markdown Table #### The Passive Flow - User sends: *"Compare AWS and GCP."* - AI starts writing a 4-paragraph history of cloud computing. - User waits 25 seconds for it to finish. - User sends: *"I meant in a table format."* - AI re-reads entire history and generates the table. - **Total Time**: 55 seconds. **Tokens Billed**: ~1,800 tokens. #### The Fast Abort Flow - User sends: *"Compare AWS and GCP."* - AI writes: *"Cloud computing has revolutionized..."* (Line 1). - **User hits STOP** (at 2 seconds). - User edits original prompt: *"Compare AWS and GCP in a 4-column Markdown table."* - User clicks Submit. - **Total Time**: 8 seconds. **Tokens Billed**: ~300 tokens. (83% Token Savings). --- ## Summary Rule of Thumb > **"Never let an AI finish a response that you already know you are going to reject."**