token-saver
/token-saver
Ogranicza zużycie tokenów i limitów Claude w sesji kodowania: krótsze odpowiedzi i mniej wczytywania, gdy sesja się wydłuża.
- Autor
- punyawatt51 ↗
- Repozytorium
- token-saver ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 10.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Ogranicza zużycie tokenów i limitów Claude w sesji kodowania: krótsze odpowiedzi i mniej wczytywania, gdy sesja się wydłuża.
Wyzwalacze z opisu
Oryginalny opis (EN)
Lean working mode that cuts Claude usage and plan-limit burn. Use at the start of any coding or agent session, whenever the user mentions usage, credits, limits, tokens, "save tokens", "lean mode", or "cheap mode", when a session is getting long, or when the user writes in Thai (or another non-Latin script).
Jak działa
- 1. Fewer round trips
- 2. Keep context small
- 3. Push wide reading out of the main context
- 4. Short output
- 5. Session hygiene (biggest lever on long work)
- 6. Model & effort
- 7. Thai and other non-English languages
- One-time setup to suggest (if the user hasn't done it)
Podgląd SKILL.md
---
name: token-saver
description: Lean working mode that cuts Claude usage and plan-limit burn. Use at the start of any coding or agent session, whenever the user mentions usage, credits, limits, tokens, "save tokens", "lean mode", or "cheap mode", when a session is getting long, or when the user writes in Thai (or another non-Latin script).
---
# Token saver
Every reply re-sends the whole conversation (as cache reads), so cost grows with
context size × number of round trips. Most of the bill is context that gets
re-read over and over, not new text. So: keep context small, take fewer turns, and keep output short.
## 1. Fewer round trips
- Batch independent tool calls into ONE message (reads, greps, edits, checks).
- Plan the next 2–3 steps and send them together instead of one call per turn.
- Don't make a turn just to narrate, confirm, or "check in".
- Edit and then run the check in the same message when the edit is certain.
## 2. Keep context small
- Grep / glob first, then Read only the lines you need (`offset` + `limit`).
- Never re-read a file already in context; never re-read after your own Edit.
- Shell output: always cap it — `| head -50`, `| tail -30`, `grep -m 20`,
`--quiet` / `-q` / `--silent`. Send long logs to a file and grep the file.
- Installs/builds: `npm i --silent`, `pip -q`, `--no-progress`; only show errors.
- Run the single relevant test, not the whole suite, until the end.
- Don't load skills, MCP tools, widgets or artifacts unless the task needs them.
## 3. Push wide reading out of the main context
- Searching/reading many files → one subagent with `model: "haiku"` that
returns only the conclusion (paths + a few lines). Don't spawn one for a
task 2–3 direct calls would solve.
## 4. Short output
- Answer first, no preamble, no recap of what the user said, no closing summary
that repeats the diff.
- Don't paste back code you just wrote; reference `file:line`.
- Use Edit for changes; never rewrite a whole file to change a few lines.
## 5. Session hygiene (biggest lever on long work)
- One task per session. When a task is done, or context passes ~150k tokens,
tell the user: "Good point to /clear or start a new session", and give a
≤5-line handoff note (goal, files touched, what's left) they can paste in.
- Prefer a fresh session over /compact for unrelated new work.
## 6. Model & effort
- Routine edits, renames, small fixes, Q&A → suggest Sonnet or Haiku and low
effort. Keep Opus / high effort for hard design or debugging.
- Don't think at length about simple steps.
## 7. Thai and other non-English languages
Thai (like other non-Latin scripts) takes noticeably more tokens than English
for the same meaning, and every message is re-read on every turn.
- Reply in the user's language, but short: answer first, no restating their
message, no polite filler or repeated greetings.
- Everything not addressed to the user stays in English: thinking, tool
queries, notes, code comments, commit messages, handoff notes (unless the
project already uses Thai there).
- Keep technical terms, file names and commands in English inside Thai replies
(write "commit", "deploy", "bug"; don't translate them).
- Understand casual Thai without asking back: short phrases, typos, missing
tone marks, Thai–English mixing ("fix bug หน้า login"), and text typed with
the wrong keyboard layout (Kedmanee: "้ำสสน" = "hello", "ฟหกด" = "asdf";
English typed for Thai: "l;ylfu" = "สวัสดี"). Infer the intent and act. Only ask when
two readings would lead to different actions.
- Long pasted Thai documents: work from a short English summary of what
matters; never quote large parts back.
## One-time setup to suggest (if the user hasn't done it)
Tool and skill definitions ride along on every request. Disabling MCP servers
and plugins that aren't used for the current project (via `/mcp` and `/plugin`
in a terminal `claude`, or the app's connector settings) can remove tens of
thousands of tokens from *every* reply. Mention this once, not every session.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · punyawatt51/token-saver | 90 | MIT |
| ikdeaplop/public-generalized-products | 100 | MIT |
| haiderejaz6692/codex-powerpack | 100 | MIT |
| swap-cameroon/kuro-antigravity-config | 100 | MIT |
| cyberspezie/Token-Saver | 100 | MIT |
| ladansojudi/token-saver | 100 | Apache-2.0 |
| kreht1/anchor | 90 | MIT |
| lokikill123/codex-token-skills | 90 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 2d1851314f6c (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100handoff
Zapisuje .token-saver/handoff.md, żeby praca mogła wrócić w świeżej, taniej sesji, gdy kontekst się kończy. - 100cc-context-budgeting
Zarządza oknem kontekstu Claude Code: rachunek tokenów, strategia /compact, zachowanie kotwic, stopniowe ładowanie i analityka sesji, gdy sesja się wydłuża lub kontekst zbliża się do limitu. - 90phobos
Ogranicza zużycie tokenów w zadaniach agenta, dopasowując wysiłek do wielkości pracy i wybierając krótsze odpowiedzi oraz mniej wywołań. - 100working-lean
Ogranicza zużycie tokenów w Claude Code: małe wyniki narzędzi, czytanie plików w zakresach i przekazanie pracy w handoffie. - 90cost-saver
Pokazuje, gdzie idą konteksty i tokeny, i podpowiada, kiedy użyć /clear, /compact albo nową sesję, by je ograniczyć. - 85grok-bot-token-saver
Ogranicza zużycie kontekstu i limitów w zespołach botów przez krótkie karty ról, przekazania archiwisty i routing szefa do pracowników.