Injecting dozens of skills, verbose rules and tool schemas into the initial prompt became the main cause of AI-agent degradation in 2026. Context bloat causes attention drift, instruction collisions and reasoning loops.
Measure it: a bloated prompt can consume several times the budget of a focused one. The rule is to keep the root rules file short, then discover the rest on demand.
The more static rules are inserted at the start of each message, the less the Transformer's attention heads can focus on the logic of the actual problem. Excessive constraints create decision paralysis and slow responses.
LCC (Local Context Compiler) normalises whitespace and removes RAG redundancies locally. The LCC repository documents exact token counts and an audit receipt for every block dropped.
Avoid describing entire directory trees, basic syntax of familiar languages or style rules that linters (Prettier, ESLint, Biome) correct deterministically. Effective rules files contain no more than 20 to 30 lines with build/test commands and non-obvious peculiarities.
3. The Progressive Disclosure Pattern
Modern architecture divides knowledge into 3 layers: 1. Minimal global rules in the root file; 2. Dynamic documentation discovery through searches (grep/ripgrep); 3. Skills and MCP servers loaded exclusively on demand when the task requires them.
4. Practical Token Optimisation Methods
Use deterministic pre-processors such as LCC (Local Context Compiler) to normalise whitespace and remove RAG redundancies, and apply session checkpoints (handoff.md) to keep conversations clean and efficient at all times.
Conclusion
Mature context engineering is not about providing everything at once, but about providing only what matters at the exact moment. Agents with lean context respond faster, make fewer mistakes and save thousands of dollars in tokens.