playbooks

home / skills / sounder25 / google-antigravity-skills-library / 08_context_pruner

08_context_pruner skill

/08_context_pruner

This skill prunes repository context to relevant files based on focus topics, reducing noise and preserving essential code.

npx playbooks add skill sounder25/google-antigravity-skills-library --skill 08_context_pruner

Review the files below or copy the command above to add this skill to your agents.

Files (3)
SKILL.md
1.4 KB
---
name: Context Window Pruner
description: Optimize information density by filtering workspace files based on relevance to the current task. Excludes high-noise files (lockfiles, assets) and prioritizes source code matching semantic queries.
version: 1.0.0
author: Antigravity Skills Library
created: 2026-01-16
leverage_score: 5/5
---

# SKILL-008: Context Window Pruner

## Overview

Large repos flood the agent's context window with noise. This skill generates a `RELEVANT_FILES.txt` list by aggressively pruning irrelevant directories and searching for files related to a specific "Focus".

## Trigger Phrases

- `focus on <topic>`
- `find relevant files for <task>`
- `prune context`

## Inputs

| Parameter | Type | Required | Default | Description |
|-----------|------|----------|---------|-------------|
| `--focus` | string | Yes | - | Keywords to filter file names/content (e.g. "gas", "auth") |
| `--max-files` | int | No | 50 | Limit output size |

## Outputs

1. `RELEVANT_FILES.txt`: List of absolute paths to read.
2. Console output summarizing token savings (estimated).

## Safety Checks

- Always respects `.gitignore`.
- Never includes binary files.
- Caps output to avoid context overflow.

## Implementation

See `prune_context.ps1`.

## Integration

```powershell
$files = .\skills\08_context_pruner\prune_context.ps1 -Focus "EVM opcode"
# Agent works only with $files
```

Overview

This skill optimizes an agent's working set by filtering workspace files to a concise, relevance-ranked list. It excludes high-noise files and directories, respects .gitignore, and produces a deterministic RELEVANT_FILES.txt for downstream reading. The result reduces token waste and focuses the agent on source files most likely to help the current task.

How this skill works

The pruner scans the repository tree while honoring .gitignore and skipping binaries, lockfiles, large assets, and other high-noise paths. It scores files by filename and content against the provided focus keywords, applies caps and safety gates, and emits a limited, ordered list of absolute file paths plus a token-savings summary. Execution is implemented as a PowerShell script for deterministic, testable behavior.

When to use it

  • Preparing an agent to work on a specific feature or bug in a large repository
  • Reducing context window usage before generating embeddings or code summaries
  • When many irrelevant assets (images, node_modules, lockfiles) crowd the workspace
  • Integrating into a pipeline to feed only high-value files to downstream models
  • Before running high-cost analysis that should avoid scanning the entire repo

Best practices

  • Provide concise, descriptive focus keywords that reflect the task intent
  • Set max-files to a practical limit (e.g., 20–100) based on model context size
  • Run the pruner early in the agent workflow and treat its output as authoritative
  • Combine filename matching with semantic content match for better precision
  • Keep the .gitignore up to date so the pruner avoids irrelevant files

Example use cases

  • Focus on 'auth' to surface authentication-related source files for a security fix
  • Prune context to 'EVM opcode' before analyzing smart-contract bytecode handling
  • Limit to 50 files when generating embeddings to stay within token budgets
  • Filter out assets and lockfiles before a refactor that targets TypeScript modules
  • Use within CI to give analysis agents a compact, deterministic input set

FAQ

How does the pruner avoid binary or large files?

It checks file types and sizes and explicitly skips binaries and large assets; it also respects .gitignore to avoid irrelevant content.

Can I rely on the ordering of RELEVANT_FILES.txt?

Yes. Files are ranked by relevance to the focus query so the top entries are highest priority for the agent.