Aggressive token optimizer and context summarizer designed for AI orchestrators dealing with massive log files, pull requests, or audit reports.
When an AI orchestrator is handling complex, multi-agent tasks or digesting massive amounts of text (e.g., repository-wide audits, massive Git diffs, or verbose build logs), its context window fills up rapidly, leading to “context bloat”, hallucination, or performance degradation.
The ask-context-janitor is a highly-constrained skill that operates as a token-optimizing subagent. Its pure focus is taking verbose inputs and ruthlessly summarizing them into minimalist, structured data points (JSON or succinct Markdown lists) that are safe to feed back into the main agent’s working memory.
This skill is intended for instances where large files or extensive outputs from parallel agents need to be digested without polluting the main contextual space.
Invoke the skill when you encounter:
.log files from build errors.ARCHITECTURAL_AUDIT.md files generated by multiple agents.git diff outputs spanning many files.The AI will output an OPTIMIZED_CONTEXT.md file stripping away boilerplate and leaving only vital information.
2026-03-07T14:32:11 [INFO] Starting build process...
2026-03-07T14:32:11 [INFO] Fetching dependencies from registry.npmjs.org
2026-03-07T14:32:12 [WARN] Unmet peer dependency for React 18, found 17
2026-03-07T14:32:15 [ERROR] Webpack compilation failed: Cannot resolve module 'crypto' in /src/utils/hash.js
2026-03-07T14:32:15 [INFO] Shutting down build process.
# Error Digest
- **Issue**: Webpack compilation failed. Missing module 'crypto'.
- **File**: `/src/utils/hash.js`
- **Warning**: Unmet peer dependency (React 18 expected, 17 found).
The agent running this skill should ideally be a smaller, faster model (e.g., Haiku or Flash) as its primary job is parsing and reduction, rather than deep logical reasoning.