Long-Conversation Anchoring and Writing Standards
You said "use PostgreSQL" at turn 1. By turn 30, the AI suddenly suggests SQLite — because the early agreement has been pushed out of the context window. This lesson covers two things: how to fight context drift in long conversations, and how to eliminate the "AI tone" from AI-generated writing. Two interactive demos — drag and click to explore.
Understand Context Drift
Context window truncation and attention decay at the tail of long texts cause AI to forget early agreements. Even 200K-Token models suffer from real attention decay toward the end of long inputs.
Set Checkpoints
After 10 turns, force a restatement of the current goal and key constraints before any critical operation — use periodic anchors to fight forgetting.
Eliminate the "AI Tone"
Replace vague instructions like "please write natural, fluent prose" with a searchable forbidden-pattern checklist and a self-review workflow.
Below is a simulated conversation window, assuming the context window can only hold the last 20 turns. Turn 1 established a hard constraint: "Use PostgreSQL." Drag the slider to add more turns and watch what happens to that agreement. Then switch to "Anchoring ON" to see the difference after the same 30 turns.
Messages with gray strikethrough have slid out of the context window — AI can no longer see them. This demo assumes a window capacity of the last 20 turns.
Restatement in a Fixed Format
Before modifying code, configs, or deploying, AI must review and restate the current goal and key constraints in a fixed format — easy to scan and confirm at a glance.
Latest Goal Takes Precedence
When the user updates the goal mid-conversation, the restatement must reflect the latest version, clearly noting the change — to prevent old and new goals from becoming mixed.
Re-read Files Before Parallel Edits
When multiple SubAgents or multiple edits touch the same file, subsequent modifications must re-read the file's current state first. Never edit based on cached or remembered old content. This is the "optimistic lock" for the multi-Agent era.
The text below was AI-generated in Chinese and is riddled with telltale "AI tone" patterns. (Note: This demo targets Chinese-language writing conventions — the sample text is intentionally kept in Chinese so the detection rules apply. Click "Run Detection" to scan each forbidden pattern per the writing-style.mdc checklist; then click each red highlight to see which rule it violates and how to fix it.)
First: "Please write in natural, fluent prose" doesn't work. What AI considers natural and what you consider natural can be completely different. You must provide a concrete list of forbidden words and forbidden sentence patterns — only then can AI execute precisely. Before delivery, search for each forbidden pattern one by one; fix each hit; note "self-check complete" when done; and also apply the forbidden patterns to any System Prompt text.
Second: Keep writing standards in their own file with alwaysApply: false — only reference it manually when writing copy or Prompts. This prevents it from polluting the context of coding conversations.
Anchors fight forgetting; checklists fight vagueness. In long conversations, periodic restatements keep agreements alive. In writing, searchable forbidden patterns keep style consistent. Both share the same logic: transform fuzzy expectations into executable actions.
Source material: Chapter 13 "Communication Standards" in rule-opensource.mdc and writing-style.mdc, open-source repository itshen/xs_vibe_rules (MIT License).