Northline Office
Menu
Free tools Law Office AI skills Videos Security Contact Book an hour
Book an hour 3D site view

AI skills / Thriftkit

Thriftkit

Spend less. Prove it. You pay for AI by the word, and the bill compounds: every message re-reads the whole conversation, so message fifty can cost many times message one. Thriftkit is three small skills that put a stop to the waste - and refuse to claim a saving they cannot prove.

Get skill - Thriftkit on GitHub Free, open source, MIT.

The proof, documented

What's in it

Three skills, one repo, nothing to pay.

token-thrift

The spending discipline. Hands your AI only the lines that matter, reminds it of what was already decided, and refuses to call anything a saving until the totals prove it.

token-thrift-audit

The receipt printer. Shows where the money went and what was saved, in plain text you can read in ten seconds.

router

The price switch. Start a work session on the cheap model or the premium model. Your choice, made on purpose.

The kits

token-thrift-audit: summary 
TOKEN SAVER - AUDIT  (2026-09-01 06:07)
------------------------------------------------
SAVED (measured):   0 across 0 job(s)
USAGE (sessions):   24,535,339,140 total tokens, 54,731 messages
  fresh input 286,416,770 | output 80,347,112 | cache-read 24,168,575,258
  top: (none)=24,058,779,840, token-saver=324,729,440,
       browser-render-audit=35,160,498
META: corpus 140 rec (~94,465 tok) | state 0 key(s)
NOTE: no savings logged yet -> measured saved = 0. Use `audit.py log`.
Captured 2026-09-01 on the machine that built this site. Note the zero: nothing was logged, so nothing is claimed. The receipt printer does not invent numbers - that is the product.
token-thrift: slice receipt 
THE RECORDED BENCHMARK RECEIPT

job:      answer one question about auth failures,
          asked against an 85 KB log file
old way:  re-send the whole file, every word billed
with slice: only the matching passages went out,
          under a hard token cap
result:   −91.91% of the tokens sent, same answer

Rule zero: nothing counts as a saving until
`job_meter.py compare` says the whole-job total fell.
Its refusal is an exit code you cannot talk past:
exit 3 · NOT A SAVING when the total went up.
Recorded in token-thrift/docs/receipts.md. The bench that produced it ships in the repo - run it yourself.