Audit agent context files for token density, directive polarity, and contradictions.
npx skills add https://github.com/broomva/skills --skill unhobble
Please help me install this Agent Skill. Skill: unhobble Source: https://github.com/broomva/skills/tree/main/skills/governance/unhobble Command: npx skills add https://github.com/broomva/skills --skill unhobble
This skill addresses the issue of over-constrained, bloated, and conflicting agent instructions that degrade performance by forcing the model to spend reasoning cycles resolving rule collisions rather than executing tasks.
Run the unhobble script on your CLAUDE.md file to generate a report identifying redundant rules and potential contradictions.
High-intent search queries and answers about installing and using this skill.
Auditing agent context involves measuring token density, directive polarity, and lexical contradictions across system prompts and governance files. This deterministic, script-based analysis identifies over-constrained instructions and enables safe deletion of redundant rules to improve agent performance.
Context engineering for Claude system prompts involves rightsizing token usage and resolving behavioral collisions by distinguishing decorative prose from load-bearing instructions. It rewrites prohibition-based rules into judgment-based heuristics to align with modern Claude-5 context engineering standards.
Rightsizing agent context involves running a deterministic script on your CLAUDE.md file to generate a report identifying redundant rules and potential contradictions. This evidence-based pruning eliminates over-constrained instructions that force the model to waste reasoning cycles resolving rule collisions.
Yes, tiktoken is used to measure token density and rules-ratio across agent governance files like CLAUDE.md and AGENTS.md. This dependency enables the deterministic, script-based analysis required to identify bloated instructions and safely prune over-constrained governance rules.
Agents waste reasoning cycles resolving rule collisions when system prompts and governance files contain over-constrained, bloated, and conflicting instructions. Lexical contradictions between directives force the model to adjudicate behavioral collisions rather than executing the intended task.
The best way to resolve behavioral collisions in AGENTS.md files is applying grounding-aware adjudication to distinguish between decorative prose and load-bearing instructions. This framework enables safe deletion of redundant rules and rewrites prohibition-based instructions into judgment-based heuristics.
georgekhananaevCompress Markdown and text into token-efficient formats with Python scripts.
AesopScottOptimizes sub-agent dispatch performance by tuning lane concurrency and fan-out parameters.
Manvendra08Reduce token consumption and context overflow in LLM interactions.
tanujsutariaValidate iOS build status, inspect Git changes, and assess session goals.
