What is Context Guard?

A safety analysis tool that evaluates JSON data in prompts to prevent unjustified certainty and reduce data leakage.

When should I use analysis, summary, or forensics?

Use analysis to identify data categories, summary to describe data structure, and forensics to investigate why a request failed or triggered a warning.

What are the requirements to run guard?

Python 3.8 or newer and the ANTHROPIC_API_KEY environment variable for token counting.

guard

Scanned

npx machina-cli add skill PJuniszewski/cook/guard --openclaw

Files (1)

SKILL.md

2.1 KB

Context Guard Skill

Epistemic safety analysis for JSON data in prompts. Prevents LLMs from reasoning with unjustified certainty when input data is incomplete.

Features

Lossless reduction - Minify, columnar transform, remove nulls
Token counting - API or heuristic fallback
Decision engine - ALLOW / SAMPLE / BLOCK
Intelligent trimming - First + last + evenly-spaced sampling
Forensic detection - Warns when specific record queries detected

Usage

When /guard is invoked, execute the guard script:

For file paths:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/guard_cmd.py" "<file_path>" [options]

For inline JSON data:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/guard_cmd.py" - [options] <<'GUARD_INPUT'
<json_data>
GUARD_INPUT

Options

Option	Description
`--mode`	analysis\|summary\|forensics (default: auto-detect)
`--force`	Bypass blocks, emit warnings only
`--allow-sampling`	Permit sampling for forensic queries
`--no-reduce`	Skip lossless reduction phase
`--budget-tokens N`	Token budget (default: 3500)
`--print-only`	Output report only, never auto-send
`--json`	Output result as JSON

Semantic Modes

Mode	Sampling	Use Case
`analysis`	Allowed	"What categories exist?", "Price range?"
`summary`	Aggressive	"Describe the data structure"
`forensics`	BLOCKED	"Why did request id=X fail?"

Output

============================================================
CONTEXT GUARD ANALYSIS
============================================================

Decision: [OK] ALLOW | [~] SAMPLE | [X] BLOCK
Mode: analysis | summary | forensics

TOKEN ANALYSIS:
  Original:     5,234 tokens
  After reduce: 4,891 tokens (-343)
  Budget:       3,500 tokens
============================================================

Requirements

Python 3.8+
ANTHROPIC_API_KEY environment variable (for token counting)

Source

git clone https://github.com/PJuniszewski/cook/blob/main/.claude/skills/guard/SKILL.mdView on GitHub

Overview

Context Guard analyzes JSON data in prompts to prevent LLMs from reasoning with unjustified certainty when input data is incomplete. It performs lossless reduction, token counting, and a decision engine that can ALLOW, SAMPLE, or BLOCK, plus forensic detection to flag risky prompts.

How This Skill Works

When invoked, the tool applies lossless reduction to the JSON (minify, columnar transform, remove nulls) and counts tokens via API or a heuristic fallback. It then runs the decision engine to produce ALLOW, SAMPLE, or BLOCK, supports intelligent trimming of the data, and flags forensic signals to warn about targeted record queries.

When to Use It

Before submitting a JSON payload to an LLM to minimize data leakage and ensure safety
When you need a quick safety audit of input data structure and completeness
When prompts might query specific records or sensitive fields and you want forensic alerts
When you want to shrink large payloads while preserving essential fields for model reasoning
When documenting or integrating prompts in regulated workflows requiring strict data governance

Quick Start

Step 1: Prepare your JSON input as a file or inline JSON
Step 2: Run guard_cmd.py with options such as --mode analysis --json
Step 3: Review the output and adjust your payload before sending to the model

Best Practices

Run with --json to get structured results and parse the decision and token data
Start in analysis mode to understand categories before moving to summary or forensic checks
Use --allow-sampling with caution when forensic style prompts are involved
Disable reduction with --no-reduce only if exact fields are required
Tune token budget with --budget-tokens to fit your model and avoid token overage

Example Use Cases

Prechecking a customer JSON payload before routing to a pricing model to avoid leakage
Redacting or trimming personal data from a large dataset before LLM ingestion
Forensic check to detect attempts to query a specific user ID or sensitive field
Summarizing a dataset structure for API documentation without exposing raw data
Auditing a multi record prompt to ensure privacy compliance in a chat assistant

Frequently Asked Questions

Add this skill to your agents