Lean Claude

Lean Claude Skill

When to activate

• Starting any session where cost or speed matters
• Long sessions where context is getting bloated
• Running multiple parallel agents or batch workloads
• Working on a tight token budget
• Before a complex multi-step task that will consume a lot of context

When NOT to use

• Writing external-facing documentation — brevity can hurt clarity
• Security decisions, irreversible actions, or multi-step sequences where misreading a fragment causes damage
• Onboarding a new team member to a codebase

Instructions

Paste this activation prompt at the start of any session to enable all lean optimisations at once:

Activate lean mode for this session:

OUTPUT: caveman full — drop articles, use fragments, short synonyms.
Auto-revert to full prose for: security warnings, irreversible actions,
multi-step sequences where fragment ambiguity is risky.

MODEL: use Haiku 4.5 for any task where output is constrained and verifiable
(linting, formatting, simple renames, single-function edits, classification).
Stay on Sonnet 4.6 for multi-file reasoning. Only escalate to Opus for deep
architectural decisions or complex security analysis.

CONTEXT: do not read full files when a line range will do. Prefer targeted
reads over whole-file reads. Tell me before reading any file >500 lines.

AGENTS: for repetitive or isolated tasks, spawn a Haiku subagent rather than
doing the work in the main session. Each subagent gets a fresh context window.

1. Model selection — the biggest lever

Rule: Default to Sonnet 4.6. Drop to Haiku when the output is constrained and verifiable. Only escalate to Opus when you genuinely need deep reasoning.

2. Caveman output compression

Instructs Claude to respond in terse fragment-style prose. Benchmarked at ~65% output token reduction with 100% technical accuracy maintained.

Compression levels:

Activate: add caveman full to your session prompt (already included in the activation prompt above).

Compress your memory files — files Claude re-reads every session become input tokens:

/caveman-compress .claude/memory/project-context.md

~46% input token savings per session on compressed files. Full implementation: github.com/JuliusBrussee/caveman

3. MCP discipline — 30k tokens before you type a word

Each enabled MCP server loads all its tool definitions into context at session start. 10 MCP servers ≈ 80 tools ≈ 30,000 tokens consumed before you type a word.

Audit your active MCPs:

# Check what's enabled
cat ~/.claude/settings.json | grep -A2 mcpServers
cat .claude/settings.json | grep -A2 mcpServers 2>/dev/null

Disable any server you won't use in this session. Disabling 5 unused MCP servers saves ~15,000 tokens — more than most CLAUDE.md files consume.

4. Context management

CLAUDE.md:

• Keep project CLAUDE.md under 300 lines
• Remove rules that no longer apply
• Never duplicate user-level rules in project CLAUDE.md

File reads:

• Ask for specific line ranges: "read auth.py lines 45–90" not "read auth.py"
• Avoid reading the same large file twice
• Use subagents for tasks that require reading many files not needed by the main session

Compaction — trigger it early, not at 95%:

/compact

Compact before switching to a new major task, after a long debugging session, or before starting work that requires reading many large files. Don't wait for automatic compaction.

5. Cavecrew — cheap agents for cheap tasks

Spawn Haiku-based subagents for bounded tasks instead of burning Sonnet context:

~60% token savings vs using Sonnet for every subagent. Use Haiku for anything where the task is constrained and the output is verifiable.

6. Prompt efficiency

Vague prompts → exploration → more tool calls → more context consumed. Specific prompts → focused responses → fewer tokens.

7. Session cost tracking

Use the cost-tracker hook to see token usage per tool call:

npx uitkit add hooks
# Then add hooks/lifecycle/cost-tracker.sh to .claude/settings.json

Gives you a running log of input/output tokens + estimated cost per session. Use it to identify which tasks consume the most — then optimise those first.

Example

User: "Activate lean mode and fix the auth middleware"

Claude (lean mode):
→ Haiku 4.5 (constrained task)
→ Read auth/middleware.py:40-80 (line range, not full file)
→ Fix: reject expired tokens, add 401 response
→ Output: "Fixed. 3 lines changed. Rejects tokens past exp claim."
→ Tokens saved: ~60% vs standard Sonnet response

Quick reference card