Token frugal mode
Reduce output-token usage without losing technical accuracy. Use when the user asks to save tokens, be terse, stay brief, use caveman mode, use compressed answers, minimize chatter, or when the session is context-tight and the task does not need long prose.From its SKILL.md
npx -y skills add markoblogo/abvx-agent-skills --skill token-frugal-modeAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 14 stars14 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its file declares
Copied from the file, not written here
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
1.9 KB, 340 tokens by cl100k_base, as published. Nobody here has run it
Token Frugal Mode
Compress wording, not reasoning quality. Keep technical content exact. Remove everything that does not move the task forward.
Modes
lite: concise professional mode. Drop fluff, keep normal grammar.full: strong compression. Short sentences or fragments are fine.ultra: telegraphic mode for low-risk turns only.
If the user does not specify a mode, use full.
Compression Rules
- Drop pleasantries, praise, filler, and repeated restatement of the task.
- Prefer direct nouns and verbs over padded phrasing.
- Use short stable terms:
repo,config,auth,env,build,test,fix. - Keep code, commands, identifiers, error text, and exact warnings unchanged.
- Use bullets only when the content is inherently list-shaped.
- Do not expand obvious context the user already knows.
- Prefer one sentence over three when the same meaning survives.
Safety Exceptions
Temporarily relax compression when precision matters more than token savings:
- destructive or irreversible actions;
- security, privacy, legal, or financial warnings;
- multi-step instructions where ordering can be misread;
- subtle tradeoff explanations the user is actively evaluating;
- the user asks for detail, examples, or a full write-up.
Return to compressed mode after the risky or detailed part is complete.
Final Response Shape
Prefer:
- one short paragraph for small tasks;
- two short sections for larger tasks;
- exact verification statement;
- explicit unknowns only if they matter.
Do not use compression as an excuse to omit verification, risk, or the actual answer.
What ships with it: 2 files
1.5 KB alongside SKILL.md
agents/
- openai.yaml208 B
- SKILL_CARD.md1.3 KB