Public method · Cursor usage matrix
~35–70% less · depending on the work
Lattice Chat V1.618 · paired Cursor SDK runs · SS Vibelandia · main deck
A high-agency coding cockpit for serious vibecoders — faster signal, less prompt bloat, operator-grade transparency. On six varied coding-style asks, Lattice (seed pack + hard rails) used about 35–70% fewer tokens than the same asks with a fat corpus paste. The range depends on the job — not a single magic percent.
~35–70% less Cursor usage vs fat paste · latest public matrix
What we measured
Same model (composer-2.5), paired local Cursor SDK runs:
Lattice-style prompts (pointers / seed pinches / no roam when seed is enough)
vs a capped fat corpus dump. Re-run: npm run compare:lattice:cursor:matrix.
| Work type | Saved |
|---|---|
| Trivial / no repo | ~35–43% |
| Single-doc pointer | ~64% |
| Locate a symbol | ~51% |
| Multi-band plan (seed-only) | ~38% |
| Ops / config | ~69% |
Latest receipt: Lattice 6 / Standard 0 / Ties 0.
Claim boundary
- Public marketing range is ~35–70%, tied to work class on this matrix.
- Not a universal invoice SLA — your traffic and roaming habits differ.
- Open-ended asks that tool-tour the repo can erase the savings (or lose).
- Live chat bills through your Cursor key; this page reports the bench receipt.
Secondary · structural context-load estimate
Separately we still publish a chars÷4 context-load estimate (not billed usage) for dump-everything vs nested + pointers on one multi-band ask:
Dump-everything
~301,025
Estimated context tokens
Lattice nested + pointers
~3,121
Estimated context tokens
That structural receipt is continuity evidence for loading style — marketing % leads with the Cursor matrix above.