How do you reduce AI token usage without lowering the quality of the result?
Prevent waste at the source. Selective Intelligence resolves intent before generation, ranks files by the current task instead of loading the repository alphabetically, reuses canonical owners, verifies once at the right boundary, and reports only the result, proof, limit, and required action.
Signals that this is the problem
- Long chats repeat the request, plan, and settled decisions before every action.
- An agent loads large folders or irrelevant files because they are nearby or sort first.
- Misunderstood intent causes whole features to be generated and discarded.
- Status replies use polished generic language but do not identify a result or proof.
What Selective Intelligence does
- Lock the wanted outcome and material constraints before spending tokens on implementation.
- Select explicit references first, then rank paths and source content by task and acceptance, expanding local dependencies and declared canonical owners.
- Set hard file, byte, and per-file budgets while excluding secrets and binary content.
- Reuse or repair the canonical owner instead of generating another implementation.
- Delete repeated setup, filler, generic praise, and technical detail that does not help the person decide or verify.
What proves improvement
- Context evidence names why every selected file matters and estimates selected and avoided tokens.
- The relevant owner wins over alphabetical filler while hard budgets and secret exclusions still hold.
- The delivered outcome and verification bar remain unchanged; only waste and correction rounds fall.
Boundary: A smaller answer is not automatically more efficient. Missing evidence, silent scope reduction, or a wrong first interpretation creates more work later and does not count as token savings.
Try this exact task
Install the public plugin, then paste this into ChatGPT or Codex with the work you want fixed:
Selective Intelligence: finish this outcome with a lean context. Read only what changes the decision, reuse existing work, avoid repeated plans and filler, and do not cut required proof.
The exact words Selective Intelligence, an unmistakable request for a named responsibility, or any correction, dissatisfaction, or failure feedback directly activates the canonical skill. The skill never widens permission to publish, spend, delete, deploy, or share.