Token saver

Ten tactics, one skill. Each one also exists as a standalone skill; full copies of all ten are in references/, so this skill works on its own. When one of them says to see another skill by name, open references/<that-name>.md.

Where the tokens go

A rule of thumb for a typical agent session, biggest first:

1. The context, re-read every turn: instructions, tool definitions, the conversation, every tool result so far. Prompt caching discounts the re-read, but it's still most of the bill.
2. Exploration: searches and file reads that happen because the agent doesn't know where to look.
3. Rework: wrong runs that finish anyway, retries, work redone after a crash.
4. Output: replies plus hidden reasoning. The priciest per token, usually the smaller total.

Model choice multiplies all four: the same work on a cheaper model costs a fraction.

Start of a session

1. Look at the meter: Claude Code /usage and /context, Codex /status, Gemini CLI /stats.
2. Pin the task down (where, observed, expected, done-when): references/pinpoint.md.
3. Plan tight? Set a budget with checkpoints: references/usage-budget.md.
4. Long job, many items or several sessions? Open a ledger: references/progress-ledger.md.
5. Big or parallel job? Plan cheap subagents with a cap: references/cheap-subagents.md.

While working

• Quiet tools, full failures: references/quiet-tools.md.
• Short answers: references/short-answers.md.
• Wrong direction? Interrupt early. Tokens already spent don't come back, but the next ones don't have to be spent. Claude Code: Esc stops the turn and keeps the work so far; a message typed while it works is queued and reaches it as soon as the running tool calls finish. Codex CLI: Enter steers mid-run, Tab queues; in the app it's Settings → General → Follow-up behavior. Gemini CLI: experimental model steering (/settings, Model Steering).
• Pick the model and the effort at the start. Switching mid-session can throw away the prompt cache (Claude Code documents this for model, effort and tool-set changes).
• Unrelated next task? Start a fresh conversation (/clear in Claude Code, /new in Codex) instead of dragging the old context along.
• Image generation is expensive: in Codex it uses limits 3–5 times faster on average than a normal turn.

After

• Did the same workflow twice? references/skillify.md.
• Runs every week? Measure it and make it cheaper: references/workflow-optimizer.md.
• Once a month: audit what loads at startup: references/context-audit.md.
• Running out every morning? Shift the 5-hour window: references/limit-window-warmup.md.

House rules to paste into AGENTS.md, CLAUDE.md or GEMINI.md

These five lines are about 150 tokens of always-loaded instructions. Habits that apply to every task work better here than as skills, which only load when a task matches them:

## Token discipline
- Before searching, confirm where (file, route, screen) and what "done" looks like; ask one short question if it's unclear.
- Use quiet flags; summarize successes in one line; show full detail only for failures, with the real exit code.
- Keep replies as short as the task allows: lead with the result, no preamble or recap.
- For work longer than about 30 minutes or across many items, keep PROGRESS.md (done / where / failed / next).
- Delegate mechanical, parallel work to a cheaper model with a spawn cap; review, don't redo.

Per tool

Universal, or specific to one model?

Almost everything here works on any model, because it's about what goes into the context and how much comes out. The exceptions:

• Budget-aware models. It's been claimed that some models can read the account's usage limits by themselves. No vendor documents that today: usage meters are shown to the user. A budget in the prompt still helps, because the model plans to it, but the agent can't enforce it without proxies. See usage-budget.
• Plan windows. The 5-hour window belongs to Claude and ChatGPT/Codex subscription plans. API billing and Gemini's quotas work differently.
• Cache timing is vendor- and billing-specific: OpenAI's GPT-5.6 and later keep cached prefixes for at least 30 minutes; Claude Code on a subscription uses a one-hour cache; API keys and usage-credit overage get five minutes.
• Price gaps between a vendor's big and small models change often. On OpenAI's current price list, GPT-5.6 Luna costs about 2% of GPT-6 Astra's per-token rate. Check the current price page before planning around a ratio.

Before you finish a heavy session

• Anything done twice is a skill now, or on the list to become one.
• The progress ledger is current, or removed if the job is done.
• If a budget was set, spend was reported against it.

Part of agent-skills · MIT · improvements welcome as issues.