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.

103 lines (79 loc) 4.36 kB
--- name: dont-repaste-entire-document description: "Reference existing document context and provide localized changes when repeating the full document is unnecessary." category: common-sense risk: safe source: self source_type: self date_added: "2026-08-26" tags: ["context-hygiene", "token-savings", "conversational-memory", "efficiency", "prompt-caching", "prompt-engineering"] tools: ["claude", "cursor", "gemini", "codex", "chatgpt"] --- # Don't Re-Paste Entire Documents (Context Hygiene) (AI Skill) ## Overview A common habit among users is re-pasting a 5,000-word document into the chat window every time they want to ask a follow-up question or tweak a single paragraph. Re-pasting entire files clogs the model's active memory window, invalidates prompt caches (making requests slower and more expensive), and risks exceeding context length limits. The **Context Hygiene Protocol** teaches users how to reference already-loaded context using anchors, section numbers, and conversational pointers. --- ## Context Bloat vs. Context Hygiene ``` ┌─────────────────────────────────────────────────────────────┐ Context Bloat vs. Hygiene Context Bloat (Re-pasting every turn): Turn 1: Paste 5,000-word document $\rightarrow$ 6,000 tokens Turn 2: Re-paste doc + 1 edit $\rightarrow$ 12,000 tokens Turn 3: Re-paste doc + 1 fix $\rightarrow$ 18,000 tokens Total: 36,000 tokens billed, high latency Context Hygiene (Pointer References): Turn 1: Paste 5,000-word document $\rightarrow$ 6,000 tokens Turn 2: "In Section 4, update point 2" $\rightarrow$ 6,100 t│ Turn 3: "Now check paragraph 3" $\rightarrow$ 6,200 t Total: 18,300 tokens (50% savings, instant responses) └─────────────────────────────────────────────────────────────┘ ``` --- ## Master Pointer Reference Prompt Templates ### Pattern 1: The Anchor-Pointer Follow-Up Use when modifying a document already present earlier in the chat: ```markdown Referring to the document you reviewed in message #1: Focus ONLY on **Section 3.2 (Payment Terms)**: - Rewrite the late fee clause to specify 2% monthly interest instead of a flat $50 fee. - Output only the revised Section 3.2 text. ``` --- ### Pattern 2: The Chapter-by-Chapter Progression Use when reviewing or drafting a multi-chapter report or ebook: ```markdown We have established the outline in our previous turn. Now let's work on **Chapter 2 ONLY**. Do NOT reprint Chapter 1 or the outline. Draft the 3 core subsections for Chapter 2 following our agreed tone. ``` --- ## Real-World Case Study ### Scenario: Refactoring a 500-Line Config / Docker Compose File #### The Wasteful Habit > **User**: Pastes the entire 500-line `docker-compose.yml` file in Turn 1. > In Turn 2, wants to add a Redis volume $\rightarrow$ Re-pastes the entire 500-line file with *"Add redis volume"*. > In Turn 3, wants to change the Postgres password $\rightarrow$ Re-pastes the entire 500-line file. #### The Context Hygiene Approach > **Turn 1**: User uploads `docker-compose.yml`. > **Turn 2**: *"Under the `redis` service defined above, add a persistent volume mount to `/data`. Show only the `redis` YAML block."* > > **AI Response**: > ```yaml > redis: > image: redis:7-alpine > ports: > - "6379:6379" > volumes: > - redis_data:/data > ``` --- ## 3 Rules of Context Hygiene 1. **Trust the session memory**: Within the same chat thread, the AI already remembers the document you pasted in Turn 1. 2. **Use specific section names**: Tell the AI: *"Look at Subsection 4B above"*. 3. **Start a fresh chat only when switching topics**: Keep related iterations in one thread, but reset when shifting to an unrelated project.