← sources

Why did a token-saving skill add ~3% to my Codex input?

source: reddit (reddit:PromptEngineering)published 2026-10-23 18:28ingested 2026-10-24 02:39

retained text

I tried recovering five current-state values from each of three synthetic project histories, roughly 20k tokens each. After switching to shorter rules and replacing a saved test log with rtk's output, mean input was 39,223 tokens. Loading the Caveman skill brought it to 40,388, another 1,165 tokens, or about 3%.

The skill instructions became part of the input. Mean provider-reported output, including reasoning, fell from 268 to about 245 tokens; the visible replies were already short. I wanted fewer tokens, then handed the model another instruction manual.

Here are the five cumulative configurations. Each row adds to the previous one:

|Configuration|Mean input tokens, rounded|
|:-|:-|
|Original request|49,621|
|\+ shorter rules|45,023|
|\+ rtk log|39,223|
|\+ Caveman skill|40,388|
|\+ TOMC history selection|28,800|

For that last step, I used TOMC to select relevant original text; I'm its author. Every configuration recovered 15/15 state values across the three histories. That small check doesn't establish general quality preservation.

This was one fresh answer per history/configuration, 15 runs total, using Codex 0.162 and gpt-6-astra (high) on October 16, 2026. Input includes cached tokens; billing savings weren't measured. Instructions changed across stages, and TOMC used raw-text selection, so this doesn't isolate each tool's effect or test compiled state.

[The same experiment as a chart](https://www.reddit.com/user/Muted-Vermicelli-180/comments/1xdl0ql/why_did_adding_a_tokensaving_skill_make_my_codex/).

Would you test longer answers next, or shorten the skill instructions and rerun this task?

  
Update, Oct 23: I shortened the skill from 1,166 to 462 cl100k tokens and reran the same three histories in fresh Codex 0.163 sessions. Full versus short averaged 40,482 versus 39,786 input tokens, and both recovered 15/15 values. Cached input differed between conditions, so this rerun doesn't establish a billing saving.

document fetched 2026-10-24 10:40 · canonical https://reddit.com/r/PromptEngineering/comments/1xdlzwy/why_did_a_tokensaving_skill_add_3_to_my_codex