Skip to content

Saving tokens

The owner's pillar (2026-11-02, #809): the same quality for fewer tokens. Before an activity, weigh whether it can be done as well for less, do the cheaper way, or suggest it when the choice is the owner's. Quality comes first: a saving that brings rework is no saving.

The practices

  1. One order per session. Every call reads the whole context again, so a long session is the main cost. Finish an order, then /clear or a new session. A UserPromptSubmit hook warns the owner and the agent once the context of a call passes 200k tokens, or CLAUDE_CONTEXT_TOKENS_REFERENCE when it is set.
  2. Filter tool output. Use --quiet, --json with --jq, tail, or Read with offset and limit: whatever enters the context is read again on every later call. A PostToolUse hook on Bash and Read warns when one output passes 10k characters.
  3. Subagents for mechanical work only, on a lighter model, with a short brief and a one-line report. What takes a few calls is done directly. A subagent that runs a scene waits for it to end and reports its verdict: its brief says so, since a subagent that stops while its scene still runs leaves the wait, the reading and the commit to the main session (owner, 2026-11-08).
  4. Shorter answers: the outcome and the next decision, with no recap of what the owner has already seen.

Both hooks are scripts/metrics/tokens, set in .claude/settings.json. They warn and never block, and whatever goes wrong in them, the session carries on.

When the context warning fires

Situation Do
Work finished or at a stopping point /wrap-up, then /clear
Mid-task with little left (a few calls: a push, a PR, a merge) Finish it in this session, then /wrap-up and /clear
Mid-task with much left (several steps, files or PRs) and the work has an issue /rollover: the handoff goes on the issue and a new session carries on
Mid-task with much left and no issue /compact
Nothing pending, all of it in a PR or an issue /clear

Suggest /rollover only when much is left (owner, 2026-11-03): it costs about a dozen calls on the long context plus a new session reading the issue again, so with a few calls left, finishing here is cheaper. /rollover is the owner's manual trigger while it is evaluated, compared with /compact on cost and rework. No /compact before /wrap-up: it spends tokens summarizing a conversation about to be cleared, and takes details the wrap-up needs.

Knowing whether it worked

Compare one week with the one before, at the Friday summary. Figures go to getnereus/internal, never here.

scripts/metrics/order summary --file metrics/orders/2026-W45.csv
scripts/metrics/tokens report --file metrics/orders/2026-W45.csv
Practice Should go down or hold Read it in
All of them, the result tokens and cost per order order summary
1, one order per session orders per session (to 1), mean and peak context tokens report
2, filter tool output big outputs, of all outputs tokens report
3, subagents tokens per agent in an order order summary
4, shorter answers output tokens per order the week's CSV
The guard: quality holds review rounds order summary

Orders differ in size, so a few of them show a direction, not a proof. A week counts as better only when the cost per order falls and the review rounds hold.