lean-context（精益环境）

Token Optimizer

Cut token usage without cutting quality. Every technique below is battle-tested in production Claude Code, OpenClaw, and agentic systems.

The 3 Token Drains (fix these first)

1. Tool output accumulation — Every file read, shell command, and MCP response appends *full output* to context permanently. A 10K-line log file stays in context for every subsequent message. This is the #1 silent drain.
2. Context compounding — The model re-reads the entire conversation on every turn. Message 50 costs more than message 5 because it re-reads 49 prior messages. Long sessions become token furnaces.
3. System prompt baseline — CLAUDE.md / AGENTS.md / system prompts load before every single request. A 5,000-token config file costs 5,000 tokens per turn, forever.

Strategy 1: Slash System Prompt Size

Target: under 500 tokens for primary config files.

# CLAUDE.md / AGENTS.md template (good — ~150 tokens)
## Rules
- TypeScript strict mode
- Test every new function
- Follow existing patterns

## Key Files
- API routes: src/api/README.md
- DB schema: docs/schema.md
- Style guide: docs/style-guide.md

Principles:

• Give the *shape* of the project, not the full docs
• Use file pointers instead of inline documentation
• Link to reference files — the model reads them only when needed
• 3-5 rules + 3-5 file pointers is the sweet spot

Strategy 2: Selective Loading (Progressive Disclosure)

Never load everything upfront. Use a three-tier system:

Implementation patterns:

File pointers over inline content:

# Bad: inline everything
## API Reference
[2000 lines of API docs]

# Good: pointer + on-demand load
## API Reference
See docs/api.md — load only when working on API endpoints.

Domain-split references:

skill/
├── SKILL.md (core workflow only)
└── references/
    ├── aws.md      # load only for AWS tasks
    ├── gcp.md      # load only for GCP tasks
    └── azure.md    # load only for Azure tasks

Conditional loading via grep patterns: For large reference files, include search patterns in SKILL.md:

# BigQuery Metrics
See references/metrics.md. Search patterns:
- Revenue queries: grep "revenue|billing" references/metrics.md
- User analytics: grep "cohort|retention|churn" references/metrics.md

Strategy 3: Compaction & Session Hygiene

When to compact vs clear:

• /compact — Context is long but thread is still relevant. Summarizes and restarts from summary.
• /clear — Switching tasks entirely. Wipes everything. Clean slate.

Rules:

• New topic = new session. No exceptions.
• Point at specific files, never "read the codebase"
• Batch related tasks in one prompt: "Fix bug, refactor, add tests" > three separate prompts
• Use scripts for data you'll read yourself — keep the agent out of the loop

Strategy 4: Efficient Tool Definitions

Tools are context too. Every tool definition loads on every request.

Principles:

• Minimal viable tool set — if a human can't tell which tool to use, the model can't either
• CLI > MCP when possible: head -20 file.log (10 tokens) vs MCP JSON response (1000+ tokens)
• Tools should return summaries, not raw dumps
• Disconnect unused MCP servers — per-request overhead is real

Tool output compression:

# Bad: full JSON dump
{"status":"success","data":{"items":[...500 lines...],"meta":{"page":1,"total":847}}}

# Good: structured summary
"847 items found. First 5: [names]. See full results in .cache/search.json"

Strategy 5: Prompt Compression Techniques

Extractive Compression (low effort, good ROI)

Select relevant sentences, discard the rest. Best for narrative documents.

# Before: 500 tokens
Customer John reported unstable internet for 3 days with video call disruptions.
Support ticket #4521 opened on 2026-04-15. Multiple attempts to reset router failed.

# After (extractive): 30 tokens
John: unstable internet 3 days, video call disruptions, router reset failed.

Selection-Based Compression (chunk-level filter)

Keep or discard entire chunks. Best for factual/citation-heavy content. Zero rewrite cost.

# Approach: filter chunks by relevance score, only pass top-k to model
relevant_chunks = [c for c in retrieved if c.score > 0.75][:3]

LLMLingua-style Token-Level Compression

Uses a small model to remove low-information tokens. Up to 20x compression with <2% quality loss.

• Use for: ICL examples, retrieved documents, long instructions
• Don't use for: code, structured data, exact citations
• Libraries: llmlingua (Python), LLMLingua2

Strategy 6: Deduplication Patterns

Across context:

• Never repeat instructions already in system prompt
• Reference previous context by turn number instead of re-stating: "as discussed in turn 3"
• Deduplicate few-shot examples — 3 diverse examples > 10 similar ones

Across sessions:

• Store shared context in files, not in conversation
• Use memory files (memory/*.md) for cross-session continuity instead of re-explaining
• Hash-check before loading: if a file hasn't changed, use cached summary

Across tools:

• Don't read the same file twice — reference previous read
• Use jq to extract specific fields instead of loading full JSON
• Pipe commands: grep ERROR log.txt | head -5 instead of cat log.txt

Strategy 7: Caching & Architecture

Prompt caching (provider-level):

• Anthropic: automatic prompt caching for repeated prefixes (system prompts, tools)
• OpenAI: cached responses for identical prompts
• Strategy: structure prompts so static content comes first, dynamic content last

Sub-agent architecture:

Main agent (clean context, ~2K tokens)
├── Sub-agent 1: deep research (uses 50K tokens, returns 1K summary)
├── Sub-agent 2: code generation (uses 30K tokens, returns diff)
└── Sub-agent 3: testing (uses 20K tokens, returns pass/fail + details)

Each sub-agent explores extensively but returns only distilled results. Main agent stays lean.

Model-tiering:

• Haiku/fast models: formatting, lookups, mechanical edits
• Mid-tier: implementation, tests, explanations
• Top-tier: architecture decisions, complex reasoning, multi-file refactors

Strategy 8: Agent-Specific Optimizations

OpenClaw:

• HEARTBEAT.md: Keep under 200 tokens. Reference external files for detailed checks.
• MEMORY.md: Curated, not raw. Review monthly. Remove stale entries.
• SOUL.md / USER.md: Load every session. Keep tight.
• Skills: Only the SKILL.md body loads on trigger. Put details in references/.
• Tool output: Use jq, head, tail, grep to scope output before it enters context.
• Write to files instead of returning large outputs in chat.

Claude Code:

• .claudeignore — Exclude node_modules, build artifacts, lock files, data files
• context-mode MCP plugin — Automatically compresses MCP tool outputs
• /compact after 20-30 messages or when switching sub-tasks
• /model to switch tiers mid-session (Haiku for lookups, Sonnet for implementation, Opus for architecture)
• MAX_THINKING_TOKENS=8000 env var to cap extended thinking budget
• Skills load on-demand only — install freely, they don't add baseline cost

General (applies to all agents):

• .gitignore pattern for AI: exclude binaries, media, large generated files
• Structured output (JSON, YAML) > unstructured prose for tool responses
• --no-ask-user / --allow-all flags reduce confirmation round-trips

Quick Audit Checklist

Run this against any AI project:

Token Audit:
□ System prompt / config files under 500 tokens?
□ Reference docs in separate files, not inline?
□ Tool outputs scoped (jq/head/grep), not raw dumps?
□ Sessions reset between topics?
□ MCP servers limited to active ones only?
□ Few-shot examples: 3 diverse > 10 similar?
□ Sub-agents used for deep exploration work?
□ Model tier matches task complexity?
□ Caching enabled for static prompt prefixes?
□ Deduplication: no repeated instructions across context layers?
□ .claudeignore / .gitignore excluding non-essential files?
□ Extended thinking budget capped for simple tasks?

Strategy ROI Matrix

References

For deeper implementation details — LLMLingua integration code, LangChain compression pipelines, TikToken measurement, sub-agent architecture patterns, semantic deduplication, and dollar savings formulas — see references/compression-deep-dive.md.