LLMBooster

A 4-step thinking framework to boost LLM output quality. Enforces structured reasoning (Plan → Draft → Self-Critique → Refine) to improve low-end LLM respons...

Security Scan

The skill is internally coherent: it provides a local, prompt-driven 4-step framework, requires no external credentials, and only reads/writes files within its skill directory — but there are minor code-quality inconsistencies and you should review the code before installing.

Purpose & Capability

Name/description (a 4-step thinking framework) matches the provided files and prompts. Required env vars/binaries are none and the code only accesses prompt files, a local JSON config schema, and a local stats file — all consistent with a local LLM prompting helper.

Instruction Scope

SKILL.md instructs the LLM to read local prompt templates and lists a CLI flow. The example /booster command runs a python3 -c snippet that imports and executes code from the skill directory; this is within the scope of a CLI-enabled skill but it means the command will execute local Python modules (not network calls). No instructions request unrelated system files, environment variables, or external endpoints.

Install Mechanism

No install spec is provided (instruction-only in registry), and all code lives in the skill package. No external downloads or package installs are present in the manifest.

Credentials

The skill requests no environment variables or external credentials. It only reads/writes files inside its own skill directory (prompts, config schema, booster_stats.json), which is appropriate for the functionality described.

Persistence & Privilege

The skill persists usage stats to booster_stats.json in the skill directory (write access). Also, the documented CLI runs Python code from the skill directory: if the skill's files are modified by an attacker, running that CLI snippet would execute arbitrary local code. The skill is NOT force-enabled (always:false) and allows normal autonomous invocation.

Assessment

What to consider before installing: - This skill is coherent with its description: it uses local prompt templates and provides CLI/status commands, and it does not ask for external credentials or network access. - It writes and reads files inside its own skill directory (prompts, config schema, booster_stats.json). That is expected for local state/stats, but be aware of persistent files being created/updated. - The SKILL.md example runs a shelled python3 -c that imports modules from the skill folder. That will execute code from the installed skill. Before running CLI commands, inspect the code in the skill directory (especially booster.py, cli_handler.py, state_manager.py) to ensure nothing unexpected was added or modified. - I noticed some code-quality inconsistencies (e.g., a malformed import line in the shown booster.py and small version metadata mismatches). These look like packaging / formatting mistakes rather than malicious intent, but they could cause runtime errors. Consider running the included unit tests in a safe environment (or reviewing the full source) before enabling. - If you install, prefer obtaining the skill from a trusted source, verify file integrity, and (if possible) run it in a sandbox or with limited permissions until you confirm behavior. - If you want, I can point out the exact files/lines with the syntax/metadata issues and suggest fixes or give guidance on how to run the tests safely.

LLMBooster Skill

A Thinking Framework, Not an Automation Tool

LLMBooster is a 4-step thinking framework that improves LLM output quality through structured reasoning. No LLM endpoint needed - the LLM follows the framework itself.

Core Philosophy

Problem with low-end LLMs: Jump to conclusions, miss details, lack self-review.

Booster solution: Enforce structured thinking process.

Plan → Draft → Self-Critique → Refine

Trigger Conditions

• User says "use booster", "booster", or "/booster"
• User requests: "detailed analysis", "in-depth analysis", "help me analyze"
• User requests: "improve quality", "detailed analysis"
• User asks for evaluation, comparison, or decision support
• User requests code review or technical documentation
• User asks complex questions (lengthy tasks, multi-step problems)

How It Works

LLM executes the framework itself, no Python calls needed:

1. LLM reads prompts/plan.md → Create structured plan
2. LLM reads prompts/draft.md → Write complete draft
3. LLM reads prompts/self_critique.md → Review issues
4. LLM reads prompts/refine.md → Polish final output

Command Handling

When user enters /booster command, execute:

cd ~/.openclaw/workspace/skills/llmbooster && python3 -c "
from config_loader import ConfigLoader
from state_manager import SkillStateManager
from cli_handler import CLICommandHandler

loader = ConfigLoader()
config = loader.load('config.schema.json')
state_mgr = SkillStateManager(config)
cli = CLICommandHandler(state_mgr)
result = cli.handle('/booster status')
print(result.message)
"

CLI Commands

Command	Description
/booster enable	Enable LLMBooster
/booster disable	Disable LLMBooster
/booster status	Show current status
/booster stats	Show usage statistics
/booster depth <1-4>	Set thinking depth
/booster help	Show help

Thinking Depth

Depth	Steps	Quality	Speed	Use Case
1	Plan	★★☆☆	Fastest	Quick analysis, brainstorm
2	Plan → Draft	★★★☆	Fast	General tasks, simple Q&A
3	+ Self-Critique	★★★★	Medium	Code review, technical docs
4	Full pipeline	★★★★★	Slowest	Important docs, complex analysis

Visual Feedback

When executing, Booster displays:

🚀 **Booster Pipeline Started**: Analyzing task...
────────────────────────────────────────
🚀 Booster [█░░░░] Step 1/4: **Plan**
✅ Plan completed (2.3s)

🚀 Booster [██░░░] Step 2/4: **Draft**
✅ Draft completed (5.1s)

🚀 Booster [███░░] Step 3/4: **Self-Critique**
✅ Self-Critique completed (1.8s)

🚀 Booster [████] Step 4/4: **Refine**
✅ Refine completed (3.2s)
────────────────────────────────────────
✅ **Booster Complete** - 4 steps, 12.4s total

Prompt Templates

All templates are in prompts/ directory:

• plan.md - Step 1: Create structured plan
• draft.md - Step 2: Write complete draft
• self_critique.md - Step 3: Review and list improvements
• refine.md - Step 4: Apply improvements

Why It Works

Low-End LLM Problem	Booster Solution
Jumps to conclusions	Plan step forces structured thinking
Misses details	Draft step requires complete coverage
No self-review	Self-Critique step finds issues
Rough output	Refine step polishes final result

Usage Statistics

/booster stats
# 📊 **Booster Statistics**
# ───────────────────────
# Status: enabled
# Thinking Depth: 4
# Tasks Processed: 5
# Last Used: 2026-03-22T09:30:00

Files

File	Purpose
SKILL.md	Skill definition + trigger conditions
README.md	Documentation
booster.py	Core module + helpers
cli_handler.py	CLI command processing
state_manager.py	State + statistics
stream_handler.py	Visual feedback
config_loader.py	Config loading
prompts/*.md	Step prompt templates