Token efficiency · concise rules
Generate token rules without dropping correctness
Select rules that reduce repeated context and unnecessary model work. The preview is compact Markdown, not a claim that the rules guarantee savings.
Purpose and audience
For operators and developers maintaining AI workflows, prompts or agent instructions. Select the policies relevant to a job, then add the generated file to that job’s instructions after checking it preserves required evidence and acceptance criteria.
How to use it
- All 16 rules are selected initially. Review the exact sentence and the output section named below each checkbox.
- Clear a checkbox to remove exactly that sentence; select it to add that sentence back.
- Review the generated Markdown preview and its fixed correctness rule.
- Copy the preview or download
AI_TOKEN_OPTIMISATION_RULES.md.
What the selections change
Each checkbox maps to one sentence in the preview: context/retrieval (minimum context, on-demand retrieval, no-repeat reads, staged hand-offs); model/tool routing (deterministic first, cheap model first, escalation, API/CLI preference, lightweight canary); or limits/state/measurement (caps, checkpoint, completed work, narration, token telemetry, cached-input measurement, marginal-value stopping). The preview keeps these headings and the correctness safeguard whether or not optional sentences are selected.
Method and limitations
This is fixed-text composition, not model execution or tokenizer measurement. “Cheapest adequate” and escalation criteria must be supplied by the workflow owner. Do not remove binding requirements or verification to reduce context.
Related: agent operating rules, context cost calculator, methodology, experiments.
Select token-efficiency rules
Select concise operating rules for reducing avoidable context and model work without dropping required evidence or verification.
Efficiency is subordinate to correctness: retrieve missing evidence rather than guessing, preserve requirements and verification, and stop when additional work has low marginal value.