Source for savetokens.tips: token-saving techniques for AI coding agents (Claude Code, Codex), published as a component datasheet. Measured figures come from per-command before-and-after token logs and refresh daily; estimates carry no log and print gray with a tilde. No number on the site is invented.

Current characterization (snapshot 2026-07-20, global scope)

[M] 79.5%	of output tokens removed by filtering command output before the agent reads it. rtk gain, 95,821 commands, 19.7M tokens saved
[E] ~50%	shorter replies with caveman mode. Deliberately lowballed target, est ~69M tokens, no before-and-after log

The guide, condensed

Tokens are metered and billed, and every word in a session gets reread on every turn, so cost compounds as the transcript grows. Everything below fights that one fact. The full version, with copy-paste blocks for CLAUDE.md and AGENTS.md, is at savetokens.tips/guide.

Audit before you optimize: /context shows what fills the window at session start, /usage shows plan spend. Then fix what the audit actually shows.

Habit	Why it works	Do this
One task per session	Every leftover token is reread on every later turn	Start fresh when the task changes
/clear between unrelated tasks	Wipes context at zero loss when nothing carries over	Cheapest move there is; use it often
/compact at checkpoints	Keeps conclusions, drops the transcript; summaries lose detail	After a fix lands, never mid-investigation
Name paths, do not paste files	A pasted file is reread every turn; an agent read happens once	"Fix the expiry check in auth middleware" beats "look at auth"
Work in bursts	Cached rereads bill about a tenth of the normal input price, and the cache expires after about five minutes	Reply while the session is hot; before a long break, /compact or /clear
Caveman mode	Reply prose is billed output; terse style cuts it roughly in half without dropping facts	Paste the copy block into your instruction file
Thin instruction file	CLAUDE.md / AGENTS.md reloads every session, so every line is a recurring cost	Cross-cutting rules only; topical detail in files loaded on demand
Prune connectors	Every MCP server ships its tool definitions into context at session start, used or not	/context shows each one's cost; disconnect the idle ones
Match the model to the task	Mechanical work does not need the strongest model at the highest effort	Route bulk work to a cheaper tier, keep the strong tier for hard problems
Subagents are not a discount	Fresh context, rereads, plus a report-back tax; total spend usually rises	Dispatch for big independent scopes to keep the main transcript short
Scripts over agent turns	Writing the script costs tokens once; every later run is free	Deterministic, repeatable work goes in a script
Filter command output	Shell dumps are the biggest silent cost; the only habit here with full before-and-after logs (figure above)	Route diffs, logs, and test runs through RTK

Measured versus estimated

The site borrows a hard convention from real datasheets: production-tested figures versus design nominals. A measured figure has a logged counterfactual (raw versus filtered tokens, per command, with a sample count) and prints green at source precision. An estimate has no counterfactual, prints gray, coarse, and with a leading tilde. Empty cells are legal; a guess never fills one.

How the figures stay live

A scheduled task refreshes the site daily from the local RTK history database:

1. export-snapshot.mjs reads history.db and writes data/snapshot.json, the single source of truth (also served publicly as the machine-readable feed).
2. apply-snapshot.mjs injects it into index.html, guide.html, llms.txt, and this README through LIVE region markers. Figures inside those markers are generated; edit the templates, never the output.
3. verify.mjs is the release gate: stale figures, a missing tilde, or a mislabeled kind fails the build.
4. daily-refresh.ps1 runs the three above, commits, and pushes; merging to main auto-deploys.

Repo map

Path	What it is
design/specimen/	The deployed site: three static pages, no build step, no dependencies
design/specimen/data/snapshot.json	Current figures, machine-readable
design/conceit.md · direction.md · constraint-contract.md	The design system: why a datasheet, and the rules the pages obey
scripts/daily-refresh.ps1	The daily pipeline runner (Windows Task Scheduler)

Run it locally

npx -y serve design/specimen        # or any static file server
node design/specimen/verify.mjs     # release gates; must pass before merging

License

MIT. The figures describe one developer's logged usage; your numbers will differ, which is rather the point of measuring.