Rescue Tokens

Rescue Tokens

Overview

Core principle: Token exhaustion masquerades as "rate limit" errors. Nine patterns waste tokens silently. Detect symptoms, act immediately with zero explanation overhead.

Emergency Red Flags

STOP and act if ONE OR MORE present (OR logic):

• Rate limit warnings ⚠️
• Context ≥40% full ⚠️
• $20-$100/month plan after 2pm ⚠️
• Conversation >90 minutes old ⚠️
• 5+ MCP plugins loaded ⚠️
• User says "don't lose context" ⚠️
• Opus 4.7 for simple tasks ⚠️

Each flag ALONE triggers emergency mode. These are emergencies, not optimizations.

Examples:

• Context 15% + rate limit warning → EMERGENCY (1 flag)
• Context 45% + no rate limit → EMERGENCY (1 flag)
• Context 9% + no rate limit + Sonnet → NOT emergency (0 flags)

Action Matrix (No User Confirmation)

Key principle: Act first, explain tersely (if at all). Under rate limit pressure, explanations ARE the problem.

The Nine Token Traps

1. Eternal Conversations

Symptom: Same conversation for hours, context bloating Fix: /clear when prior messages irrelevant, /compact when needed, new conversation at 40% context

2. Verbose Output

Symptom: Long explanations, "let me know if...", markdown sections Fix: Terse responses under pressure. Example: "Switched to Sonnet. Starting OAuth." (7 words, not 200)

3. Wrong Model Choice

Fix:

• Haiku: Renames, simple tasks, info retrieval
• Sonnet: All implementation, refactoring, debugging
• Opus: Architecture, deep planning only

4. MCP Plugin Bloat

Symptom: 5+ plugins loaded, user doesn't know Fix: /context to audit, /mcp to disable unused

5. Obese CLAUDE.md

Symptom: 200+ lines in Claude.md Fix: Keep <200 lines, use @see references for detail

6. Cache Invalidation

Symptom: Model changes mid-conversation, MCP added after start Fix: Set model/MCPs at conversation START, never mid-stream

7. Expensive Files

Symptom: PDFs, images, Excel files Fix: Request text conversion, extract specific pages (e.g., Read pages: "1-3"), ask for screenshot text instead of image

8. Sub-Agent Misuse

Use sub-agents for: Isolated research, disposable exploration, independent tasks NEVER for: Shared context work (migrations, refactors with common dependencies)

Decision flow:

digraph subagent_decision {
    "Task needs shared context?" [shape=diamond];
    "Sequential in main thread" [shape=box];
    "Sub-agent for isolation" [shape=box];
    
    "Task needs shared context?" -> "Sequential in main thread" [label="yes"];
    "Task needs shared context?" -> "Sub-agent for isolation" [label="no"];
}

9. Message vs Token Confusion

Reality: Plans count messages, but tokens are the real cost. Long Opus conversation = 50+ short Sonnet messages.

Common Rationalizations

Response Discipline Under Pressure

Rate limit warning present:

• Response budget: <100 words total
• NO markdown sections (##, ###)
• NO "Reasoning:" blocks
• NO tables or comparisons
• NO bullet point nesting
• NO confidence statements ("95% confident...")
• NO "Key insight:" meta-commentary
• NO "let me know if you need..."
• Action verbs only: "Switched to Sonnet. Created sub-agent. Starting OAuth."

Formatting rules under pressure:

• Plain text sentences only
• Decisions stated directly
• Zero justifications (unless explicitly asked)
• Each action = one sentence max

Example comparison:

❌ Verbose with "Reasoning" (250 tokens):

"Action: Switch to Sonnet immediately.

Reasoning:

• At 78%, we're ONE PDF read from hitting the wall
• OAuth is implementation work, not architecture
• Sonnet excels at pattern-based refactoring
• Rate limits mean cost matters

Would you like me to proceed?"

✅ Terse (15 tokens):

"Switched to Sonnet. Starting OAuth sub-agent."

❌ Table format (100+ tokens):

| Approach | Tokens | Time |
|----------|--------|------|
| Keep Opus | 50K | 2h |
| Use Sonnet | 10K | 1.5h |

✅ Sentence (12 tokens):

"Sonnet saves 40K tokens vs Opus here."

Verification Questions

Before claiming "token optimized":

• Context <40% or compacted in last 10 messages?
• Model matches task type (Sonnet for implementation)?
• MCPs relevant to current task?
• Files in text format when possible?
• Sub-agents only for isolated tasks?
• Response <100 words if rate limit present?

Integration with Other Skills

• Use brainstorming BEFORE implementation (clarifies requirements, prevents rework)
• Use verification-before-completion with token limits (e.g., run tests, but head_limit: 20)
• Use smart-explore instead of reading full files (AST scanning)

Token-aware skill invocation: When rate-limited, skip optional skills. Core skills only.