token-saviour
Skill2 starsChanged 3 months agoMITView on GitHub
Route each sub-task to the cheapest adequate tool across a 4-layer routing table (serena / rtk / caveman / Ponytail), cutting overall token spend by ~70%. Use when an agent loop's context is ballooning, when choosing between heavy and light tooling for a sub-task, or when budgeting a long-running job.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- vagkaratzas/skillstheir own version, first seen Oct 10, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- vagkaratzas/skills★ 4same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository7 other files it gives its agents.
- audit-swarmSkill
- cavemanSkill
- grill-meSkill
- prism-reviewSkill
- prompt-forgeSkill
- slide-smithSkill
- token-wardenSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 18
Before spinning up an expensive coding agent for a small, well-scoped edit.
- Distinctive: line 19
When a conversation history exceeds a token threshold and must be trimmed.
- Distinctive: line 20
When deciding which model/agent should own a sub-task.
- Distinctive: line 21
When a task can be done by a formatter, a linter, or a grep instead of a chat.
- Distinctive: line 41
Run Layer 3 on everything that crosses into Layer 4.
- Distinctive: line 46
After the sub-task, report the delta vs.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Token Saviour
Token discipline is a routing problem, not a willpower problem. token-saviour picks the lightest tool that can still do the job, then compresses what must be sent. It is the operational layer beneath token-warden's budget accounting.
When to use
- Before spinning up an expensive coding agent for a small, well-scoped edit.
- When a conversation history exceeds a token threshold and must be trimmed.
- When deciding which model/agent should own a sub-task.
- When a task can be done by a formatter, a linter, or a grep instead of a chat.
Workflow
Step 1 — Classify the sub-task
Tag it on two axes: token-cost sensitivity (how much context it drags) and adequacy floor (minimum capability needed). A rename is high-sensitivity, low-floor; an architecture decision is high-sensitivity, high-floor.
Step 2 — Consult the 4-layer routing table
See references/routing-table.md for the full matrix. In short:
- Layer 1 — serena: structural/semantic code ops (symbol rename, find-usages) that a language server does cheaper than an LLM.
- Layer 2 — rtk (regex/toolkit): deterministic text transforms, search, bulk edits a script handles without a model.
- Layer 3 — caveman: compress any context before it reaches a model.
- Layer 4 — Ponytail (heavy agent): only when lower layers cannot meet the adequacy floor.
Step 3 — Compress inputs, cap history
Run Layer 3 on everything that crosses into Layer 4. Cap rolled-up history to the last N meaningful turns; archive the rest to a retrievable note, not the live context.
Step 4 — Measure and report
After the sub-task, report the delta vs. the naive "send everything to the heavy agent" baseline. A good call shows savings; a bad call shows where the floor was missed.
Anti-patterns (never do these)
- Routing everything to Layer 4 "just to be safe" — that is the bug it fixes.
- Compressing load-bearing context to hit a number — adequacy first, thrift second.
- Skipping the measurement — unseen savings aren't savings.
- Treating the table as fixed — re-tune thresholds per project.
Output contract
Return a routing decision: chosen layer(s), the estimated token delta vs. baseline, and one line of rationale. If no layer meets the adequacy floor, say so explicitly and escalate rather than guessing.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.