Spec index cut AI reading by about 97.2% in developer's test

The setup grew out of a problem any team using coding assistants knows: handing the whole spec to the AI every time bloats the reading volume, but trimming it too aggressively drops the conditions that matter. The author's answer is to separate the specification into per-component HTML documents and give the same specs two doors — a mind-map view for people to see where things are, and an external program that pulls out only the relevant lines for the AI.

The compression figures come with firm caveats. The token counts are estimates from UTF-8 byte counts divided by four, not model-specific tokenizers or billing records. The 約97.2% came from reading an 803-line index file; a code-search extraction returned 777 tokens against 12,444 for roughly 約93.8% compression, and a screen-check diagnostic returned 339 against 2,024 for about 約83.3%. The first two returned all matches with no count-based cutoffs, and the third involved 8,096 bytes with zero findings or page errors.

The design is not a semantic search: candidate selection uses keywords and weighting, and past-document judgment leans on file names and locations. Whether the savings hold on a wider, cross-cutting change is likely to depend on how often the AI still has to open the original spec and code to verify conditions — which is precisely the check the author says should never be skipped. A follow-up piece is planned on the token-measurement program itself.