Agent-Skills-for-Context-Engineering

**Judgement:** High-quality, production-aware skill for context compression and structured summarization, especially strong for coding/debugging agents and large codebases. Best suited for advanced agent builders designing memory/compression layers rather than end users looking for a simple summarizer. **What it does well** - Reframes optimization around **tokens-per-task**, emphasizing re-fetching and recovery costs instead of raw compression ratio. - Provides a robust **anchored iterative summarization** pattern with explicit sections (intent, files modified, decisions, state, next steps) that directly target common failure modes in coding agents. - Clearly compares **anchored iterative vs regenerative vs opaque** compression, including compression ratios and quality trade-offs. - Offers practical guidance on **compression triggers** (threshold-based, sliding window, importance-based, task-boundary) and recommends sliding window + structured summaries for most coding scenarios. - Introduces **probe-based evaluation** (recall, artifact, continuation, decision probes) for measuring functional quality rather than surface similarity. **Key risks and limitations** - **Artifact trail weakness:** The skill explicitly notes that tracking created/modified/read files remains a weak point across methods; reliable artifact tracking may require **separate indices or explicit file-state tracking** beyond summarization. - **Complexity of integration:** Effective use assumes a reasonably sophisticated agent scaffolding (summary memory, compression triggers, probes, artifact index), so it is not plug-and-play. - **Information loss under misconfiguration:** Poorly tuned triggers or over-aggressive compression (especially opaque methods) can cause silent loss of critical details (file paths, error messages, decisions), increasing re-fetching and hallucination risk. - **Coding-focused bias:** Guidance and examples are optimized for coding/debugging agents; other domains may need adapted summary structures. **Recommended scenarios** - Long-running **coding or debugging agents** where context windows are frequently exceeded and file/decision tracking is critical. - Large **codebase analysis or migration** workflows (multi-million-token systems) where research, planning, and implementation must be compressed into durable specs. - Designing **memory/compression subsystems** for general-purpose agents that need structured, phase-aware summaries instead of ad-hoc truncation. - Building **evaluation frameworks** to compare different compression strategies or vendors, using probe-based tests for functional quality. Use this skill when you need a detailed, methodical approach to context compression and are willing to invest engineering effort into structured summaries, artifact tracking, and evaluation. It is less appropriate if you need a simple, one-shot summarizer with minimal integration work.