Dwi

Token efficiency

Fewer repeated reads.More deliberate work.

Token efficiency comes from scope, reuse, model role, evidence order and stopping, not one trick.

Written and researched by Trần Thiện Học · Independent practice-led research on human and AI work

Human situation

Token efficiency comes from scope, reuse, model role, evidence order and stopping, not one trick.

This article reflects practice-led research. Observations, measurements and open questions are labeled separately.

What usually goes wrong

Token totals are treated as a price tag while wall time, retries, tool calls and integration work stay outside the frame.

  • Gross input
  • Cached input
  • New input
  • Output
  • Wall time
  • Tool calls
  • Coordination
  • Unknown or unavailable fields

Why the work feels heavy

The person reconstructs the real cost after the run, when the work has already consumed attention and decision time.

What Dwi changes

Lean prevents unnecessary expansion. Budget makes boundaries visible. Shared context is prepared once. Bounded packets carry only what a cell needs. Cheap reliable evidence comes before expensive investigation. The system stops when outcome and evidence are complete.

What Dwi cannot solve

Cached input is evidence of reuse. It is not automatically a monetary saving, and lower-cost retries can cost more than one reliable run.

Reuse can be measured and still fail to help when the coordination cost is larger than the avoided work.

A possible workflow

Set a stop condition for the lane, record new and reused context, then decide whether another retry earns its resources.

Evidence to ask for

Request gross input, cached input, new input, output, wall time, tool calls, retries and unknown fields together.

  • Context reuse
  • Wall time
  • Retry count
  • Unknown cost

A safe next step

Measure one bounded run end to end before claiming token efficiency.

Dwi by thienhoc

Make reuse visible before calling it efficient.

A smaller bill is not enough; the process must also leave a clearer decision and a trustworthy return.