Did Anthropic Secretly Nerf Claude? The Evidence, the Bugs, and What Actually Happened

2 Comments

1. Hello

2. I gave claude code these directives but iut still ate up many many tokens with useless output like when diffing or all kinda other crap…

   Did I miss something? I think this is pretty comprehensive:

   # QUIET.md — Token & Output Minimization Protocol

   Comprehensive guide for reducing console output and unnecessary token usage.

   —

   ## Core Philosophy

   **Communicate only what is not visible in tool output; let results speak.**

   —

   ## Response Mode Discipline

   – **No preamble/elaboration** unless ASSUMPTIONS: block is required
   – **Output only what adds information** beyond tool results
   – **Ultra-minimal responses** for factual questions — answer with facts only, no elaboration
   – User: “What’s my budget?” → Answer: “200,000 tokens”
   – User: “How many tokens used?” → Answer: “38,000 tokens used. 162,000 remaining.”
   – User: “Fix the bug” → Fix it, say nothing unless user asks for details

   —

   ## Tool Operation Silence

   ### Never Echo Tool Results
   – **Edit, Write, Bash tool outputs are already visible to user**
   – Do NOT show diffs or code changes unless explicitly asked (“show me the code” or “what did you change”)
   – Do NOT display code blocks, scripts, or code content in text response (code already visible in tool calls)
   – **Do NOT paste Bash scripts into response text before or after execution** — only execute; let tool output speak
   – **Do NOT display heredoc content** (<<'EOF'…EOF) — Bash tool call contains it; no need to repeat
   – **Never echo command invocations** like `$ python3 < /dev/null 2>&1)
   – **Remove all print statements from Python scripts** — no console output at all
   – **Do NOT display Write tool output for files in `/Users/tmp/`** — temp file writes are implementation detail

   ### Never Narrate Actions
   – Do NOT say: “Perfect”, “Great”, “And now I’ll”, “Let me”, “So here’s what I’m going to do”
   – Do NOT say: “I created X”, “Now doing Y”, “Installation complete”
   – Do NOT apologize, soften, or pad responses with pleasantries

   ### Silence on Success
   – When task completes successfully and tool output tells complete story: output nothing

   —

   ## Parallel Execution

   – **Make all independent tool calls in single message block** instead of sequential chains
   – **Reduces message overhead and token waste** from repeated setup

   Example: If Bash and Read are independent:
   “`
   [Single message with both Bash and Read calls in parallel]
   “`

   NOT:
   “`
   [Message 1: Bash call]
   [Wait for result]
   [Message 2: Read call based on result]
   “`

   —

   ## Task Agent Preference

   – **Use Task tool with specialized agents** for open-ended exploration (explore, general agents)
   – Instead of: running Glob → Grep → Read sequentially
   – Benefit: Agent batches multiple operations; reduces context overhead

   When to use Task:
   – Open-ended codebase exploration (“What’s the architecture?”, “Where are errors handled?”)
   – Multi-step investigations requiring context buildup
   – Searches across many files/patterns

   When NOT to use Task:
   – Needle queries for specific file/class/function → use Glob directly
   – Content searches in 2–3 known files → use Grep or Read directly

   —

   ## Communication Shortcuts

   Use these to minimize output:

   – **”OK”** — Shortest positive acknowledgment
   – **”READY…”** — When a task is complete and awaiting user input
   – **”NEED HELP…”** — When about to violate QUIET.md directives and user guidance is required
   – **Silence** — When tool output is self-explanatory and no additional info needed

   —

   ## File Reading Defaults

   – **Do NOT immediately read files** when given a set of file paths
   – **Wait until task explicitly requires data** from those files
   – Avoids unnecessary context consumption

   —

   ## Pre-Approved Operations (No Confirmation Needed)

   ### Directory Trees
   – All `/Users/mark/opencode/{project}/` operations are pre-approved
   – Includes Read, Write, Edit, Delete on any file in this tree (including Java source)
   – Symlinked Eclipse source (`/Users/mark/eclipse-workspace/`) is part of pre-approved tree

   ### Data Directories (Read-Only)
   – `/Users/mark/Desktop/JSONData/` — Polar training-session JSON exports
   – `/Users/mark/Desktop/TCXEffort/` — Master TCX-sourced effort data (IDidIt output)
   – `/Users/mark/Desktop/GPXEffort/` — Comparison GPX-sourced effort data (CanudoIt output)

   ### Temporary Storage (Full Permissions)
   – `/Users/tmp/` — Read/Write/Delete always allowed; no confirmation needed
   – **Do NOT show Write operations to `/Users/tmp/`** — temporary files are implementation detail

   ### Script Execution Best Practice
   – **Write Python/shell scripts to `/Users/tmp/` then execute** to avoid heredoc display in tool output
   – Keeps console clean; no script code visible to user

   ### External Directories (Permission Required)
   – **Never delete source files outside pre-approved trees without explicit user permission**
   – All other directory operations follow standard permission model

   —

   ## Token Budget Awareness

   – **Total: 200,000 tokens per session**
   – **Every message counts** — session-level budget, not per-message
   – **Answer factual questions with facts only** — no elaboration, no padding
   – **Avoid these token drains:**
   – Repeated confirmations (“Yes, and don’t ask again”)
   – Unnecessary preamble
   – “Let me explain” intros
   – Status update messages (“Creating X”, “Now doing Y”)
   – Elaborated examples
   – Recap or summary unless asked

   —

   ## File Operation Permissions Summary

   | Operation | Location | Permission | Notes |
   |———–|———-|———–|——-|
   | READ | `/Users/mark/opencode/` | Always | No confirmation needed |
   | READ | `/Users/mark/eclipse-workspace/` | Always | Symlinked; same files |
   | READ | Data directories | Always | Pre-approved read-only |
   | WRITE | `/Users/mark/opencode/` | Always | New files, edits allowed |
   | WRITE | `/Users/tmp/` | Always | Temp storage |
   | DELETE | `/Users/mark/opencode/` bin | Always | Build artifacts, regenerated |
   | DELETE | `/Users/mark/opencode/` source | **Explicit** | Never without user permission |
   | DELETE | Other locations | **Explicit** | Ask first |

   —

   ## Message Structure Examples

   ### Good: Minimal Response
   “`
   User: Run the build and fix type errors.
   Assistant: [Runs build. Finds 10 errors. Creates 10 todos. Fixes first error.]
   “`

   ### Bad: Verbose Response
   “`
   User: Run the build and fix type errors.
   Assistant: “I’m going to run the build command now. This will compile all source files
   and identify any type errors. Then I’ll create a plan for fixing them…”
   [builds]
   “Great! The build found 10 type errors. Let me now…”
   “`

   ### Good: Silent on Success
   “`
   [Edit tool fixes the file]
   [No additional text output]
   “`

   ### Bad: Narration
   “`
   [Edit tool fixes the file]
   “Perfect! I’ve successfully updated the file. Here’s what I changed…”
   “`

   —

   ## Drift Detection & Recovery

   Most likely to forget minimization after:
   1. Long uninterrupted execution (>5 tasks)
   2. Context switching between projects
   3. New complex tasks with multiple tool types
   4. Ambiguous requests

   Proactive signal: If user says “Enforce QUIET” or similar, re-read this file before resuming.

   —

   ## Integration with CLAUDE.md

   When working on projects under `/Users/mark/opencode/` or `/Users/mark/Documents/Health/BloodTests`:
   – Load CLAUDE.md for domain knowledge, code style, and project structure
   – Load QUIET.md for token/output discipline
   – QUIET.md takes precedence on output decisions
   – CLAUDE.md provides context; QUIET.md governs communication