context hygiene strategist

The problem

Long agent sessions suffer from context bloat, leading to high API costs and a documented 23 to 31 percent drop in instruction-following accuracy. When an agent loses the plot or repeats itself, it is usually because the context window is crowded with low-value tool outputs and dead-end reasoning.

What it does

• Audits session context to identify verbatim tool dumps, redundant facts, and abandoned reasoning branches.
• Implements a token budget protocol with tripwires at 50, 75, and 90 percent of capacity.
• Compacts sessions into a high-density working-state block that preserves decisions and constraints while discarding fluff.
• Enforces retrieval discipline by requesting specific document slices rather than full-file ingests.
• Calculates phase-based costs to identify whether research, execution, or verification is draining the budget.

Why this beats prompting it yourself

General prompts fail to maintain hygiene over long durations, often resulting in "memory" that is just a messy log of every interaction. This skill provides a structured economic framework and a specific compaction playbook that forces the agent to treat tokens as a finite financial resource, maintaining model reasoning quality at scale.

Use cases

• Managing multi-hour coding sessions where tool outputs and linter logs threaten to drown out the original requirements.
• Running large-scale research tasks where full-document retrieval would otherwise blow the token budget.
• Optimizing production agent costs by replacing verbose narrations with terse, state-based progress tracking.
• Recovering "lost" agents that have become sluggish or repetitive due to context saturation.

Known limitations

If the LLM provider does not expose real-time token counts, the strategist relies on message size estimation. It requires the agent to have the ability to summarize and replace its own history.