· · · · · ·

🦴 paleo

Token-saving skills for LLM agents — cut output & context tokens without choking the model.

Why · See it · Features · Skills · Combos · Quick Start · Benchmarks · Comparison · Tips · Install

Why paleo?

Tokens cost money and latency. Every trimmed token means faster, cheaper inference.
One-size-fits-all prompting fails. Sometimes you want terse output, sometimes a hard budget, sometimes just no fluff. paleo gives each as a separate, well-scoped skill.
Skills stay minimal. Every SKILL.md is written terse on purpose — loading one costs less context than a long prompt.

See it (Before / After)

paleo compresses delivery, not meaning. Code, commands, and technical terms stay byte-exact.

Normal agent	🦴 paleo
"The re-render happens because you create a new object literal on every render. That inline object is a fresh reference each time, so React sees a changed prop and re-renders. Wrap it in `useMemo` to keep a stable reference."	"New object each render → new ref → re-render. Wrap in `useMemo`. Stable ref = no re-render."
"To authenticate requests, add middleware that checks the token on each request and returns 401 if it is missing or expired."	"Add auth middleware. Check token per request. 401 if missing/expired."

[!NOTE] paleo keeps technical accuracy at 100% — it drops filler, not facts.

Features

Modular & composable — load one skill or all seven, mix per task.
Output + context savings — ~50–70% fewer output tokens (median ~54% on a 6-task sample — see BENCHMARK.md), plus proactive context trimming.
Auto-detect — paleo-auto watches session state and enables the right skills automatically. No thinking required.
Production-safe — compresses output and context only; never rewrites your code.
Hard token budget — paleo-budget caps spend and summarizes the tail.
Cross-agent — open Agent Skills standard: Claude Code, Codex, Gemini CLI, Qwen Code, OpenCode, Cursor, GitHub Copilot, Cline, Windsurf + 40+ agents via npx skills add.
Zero-setup triggers — plain English phrases, no slash commands to register.
Low overhead — each SKILL.md is intentionally terse, so loading stays cheap.
Open & extensible — drop in your own token-saving skills.

Skills

Skill	What it does	Trigger example
`paleo`	Terse output mode — cut output tokens ~50–70%, keep code/terms exact.	`paleo mode` · `be brief` · `save tokens`
`paleo-auto`	🆕 Zero-touch auto-detection — watches session & enables the right skills automatically.	`paleo-auto` · `auto paleo` · `enable auto-save`
`paleo-budget`	Hard token budget per task — cap spend, summarize if exceeded.	`budget 2000` · `stay under 2000 tokens`
`paleo-trim-context`	Proactively trim/summarize context to save tokens without losing task state.	`trim context` (auto on long sessions)
`paleo-converse`	Condense old chat turns + merge duplicate messages; keep last N verbatim.	`condense chat` · `compress conversation` · `paleo-converse N=8`
`paleo-summary`	Tight intisari of bulky tool output / logs / diffs / dumps.	`tldr` · `condense this` · `summarize output`
`paleo-json`	Minify & compact structured/JSON output, stay parseable.	`compact json` · `minify`

Recommended combos

paleo is the base — keep it on. Layer the rest by what you're doing:

Situation	Combo	Why
Daily driver (long / chatty sessions)	`paleo` + `paleo-trim-context`	Base + automatic context hygiene.
History piling up	+ `paleo-converse`	Condense + merge duplicate turns once a session gets long.
Debugging / bulky tool output	`paleo` + `paleo-summary` + `paleo-json`	Logs → intisari; JSON → minified.
Tight cost / hard limit	`paleo` + `paleo-budget` (+ `paleo-trim-context`)	Hard ceiling + shrink context first.
Max savings (all on)	all seven skills	Overkill daily, but safe for extreme thrift.

[!TIP] You rarely need every skill at once. paleo + paleo-auto is the default for most users — auto-detect handles the rest. For manual control: paleo + paleo-trim-context is the baseline; add paleo-converse for messy chats, paleo-summary / paleo-json for heavy tool output, and paleo-budget only when a hard cap is required.

Quick Start

# 1. Clone the collection
git clone https://github.com/mocasus/paleo.git

# 2. Claude Code — one plugin bundles all 7 skills
claude plugin marketplace add https://github.com/mocasus/paleo
claude plugin install paleo@paleo

# 3. Any agent via the open Agent Skills registry (installs to 40+ clients)
npx skills add mocasus/paleo

Then just talk to your agent — no command to register:

paleo mode · save tokens · budget 2000 · trim context

Benchmarks

Real, reproducible numbers — not hand-waved claims.

Model	Tasks	Median output savings	Mean
`claude-sonnet-4.5`	6	53.8%	45.1%

Full method, per-task table, and the runnable harness are in BENCHMARK.md. Rerun on your own stack:

export IDROUTER_API_KEY=your_key
python3 bench/benchmark.py --model claude-sonnet-4.5

[!TIP] Savings are task-dependent: biggest on verbose generative work (code, walkthroughs, comparisons — 54–79%), smaller on already-compact factual answers. paleo also cuts context tokens via paleo-trim-context, a layer a terse-persona prompt cannot reach.

Comparison

paleo is often compared with two other token-saving approachesa terse-persona system prompt and Ponytail (a code-reuse coding skill). Here is how they differ.

Dimension	🦴 paleo	Terse-persona prompt	Ponytail
Form	6 composable skills	Single system prompt (persona)	Single coding skill / workflow
What it targets	Output tokens + context + conversation turns	Output tokens only	Volume of code the agent writes (+ MCP caching)
Granularity	Per-task, mix & match	One mode	One workflow
Touches your code	❌ No (output/context only)	❌ No	⚠️ Yes — refactors / reuses code
Context & reasoning savings	✅ `paleo-trim-context`	❌ None	◑ Partial (caching)
Hard budget	✅ `paleo-budget`	❌	❌
Cross-agent	✅ 40+ agents (open standard)	➖ Portable prompt, but monolithic	➖ Claude Code skill
Activation	Plain phrases	Edit system prompt	Install + invoke skill
Reasoning-model safe	✅ Never compresses thought	❌ Can raise tokens (e.g. +3% on Opus)	➖
Known risk	None (output-only)	Can fight "expand" heuristics; may raise tokens on reasoning models	Refactor can change behavior
Open benchmark	✅ Reproducible harness	❌ Claim only	❌ Claim only

[!TIP] They're complementary, not rivals. Ponytail cuts the code you have to write; paleo cuts the tokens in the conversation. Terse-persona prompts proved a terse prompt helps output — paleo takes that same idea and makes it modular, adds context-trimming and a hard budget, and drops the persona gimmick. Use Ponytail for code-heavy work and paleo for chatty, long sessions.

FAQ

Gak lebih jelek — malah sering lebih bagus. paleo bukan potong buta, dia buang redundansi (ulang-ulang, filler, boilerplate hasil tool, whitespace), bukan info esensial. Constraint, error, code, keputusan tetap utuh. Prompt lo gak diapa-apain — yang di-trim itu context kerja agent (hasil tool berulang, riwayat convo). Context bersih = model fokus ke signal, bukan lost track karena noise. Benchmark: median 53.8% token turun, kualitas task gak drop. Trade-off jujur: budget ekstrem (token cap rendah banget) bisa turun kualitas, tapi setting optional. Intinya: lebih murah & cepet, kualitas tetap.

paleo gak pernah sentuh system prompt. System prompt = aturan main, utuh 100%. Instruksi yang define behavior gak diapa-apain — paleo jalan setelah system prompt ke-load, cuma kerja di context dinamis. Yang di-trim itu context kerja (tool output berulang, convo kepanjangan), bukan instruksi. System prompt kecil dibanding noise yang numpuk dari tool output. Plus bisa whitelist bagian yang mau dijaga. Efeknya kebalik: context bersih bikin model lebih patuh ke system prompt. Aturan main tetep nempel.

Gak. paleo melindungi code block, structured output (JSON/table), dan error message secara default. Yang di-compress cuma prose bertele-tele & tool output redundant. Kalo masih ragu, bisa whitelist file/section tertentu biar 100% gak ke-trim.

Enggak. paleo murni teknik prompt/context — gak ada server, gak ada API call, gak ada dependensi luar. Skill-nya tinggal di-load ke agent lo, jalan di lokal. No overengineering.

Kelihatan, terutama di session panjang & agent loop. Tiap token yang gak dikirim = gak dibayar. Benchmark kita median 53.8% turun di context + output. Di agent yang muter 20+ tool call, itu selisih gede per run.

Iya. paleo model-agnostic — kerja di level prompt & context, bukan di model tertentu. Claude, GPT, Gemini, GLM, Qwen, lokal — semua bisa. Sifatnya instruksi, bukan fine-tune.

Compaction bawaan itu generic & reaktif (baru jalan pas context mau penuh, sering blind truncation). paleo proaktif + selektif: jaga info esensial, buang redundansi, configurable, ada safety net (whitelist). Plus paleo juga ngurus output verbosity & tool-result summarization, bukan cuma convo history.

Trigger pakai natural language (skip preamble, ringkas output, trim context) — gak perlu slash command. Mau matiin? Tinggal gak dipanggil, atau cabut skill dari agent. No global lock-in.

Tips & Triggers

paleo activates from natural-language triggers — no slash command to register. Type the trigger, the skill loads and applies.

paleo — terse output

On: paleo mode · be brief · terse · compress output · save tokens
Level: paleo full (default) · paleo lite · paleo ultra
Off: stop paleo · normal mode

paleo-budget — token cap

On: budget 2000 · stay under 2000 tokens · token limit
Off: no budget · unlimited

paleo-trim-context — auto on long sessions; trim context to force. paleo-auto — paleo-auto · auto paleo — off: disable paleo-auto · manual paleo paleo-converse — condense chat · compress conversation · paleo-converse N=8 paleo-summary — tldr · condense this · summarize output paleo-json — compact json · minify

Combo: paleo + paleo-budget = max savings. Add paleo-trim-context on long sessions, paleo-converse on chatty ones, paleo-summary for bulky tool output.

Installation

claude plugin marketplace add https://github.com/mocasus/paleo
claude plugin install paleo@paleo

npx skills add mocasus/paleo

mkdir -p ~/.gemini/skills && cp -r skills/* ~/.gemini/skills/

hermes skills install mocasus/paleo
# Or copy skills manually
cp -r skills/paleo* ~/.hermes/skills/

npx skills add mocasus/paleo

npx skills add mocasus/paleo

npx skills add mocasus/paleo

git clone https://github.com/mocasus/paleo.git
# copy skills/paleo*/ into your agent's skills directory

All 7 skills load automatically — paleo, paleo-trim-context, paleo-auto, paleo-budget, paleo-converse, paleo-summary, paleo-json.

Full per-agent steps in INSTALL.md. See real compression numbers in BENCHMARK.md.

Hermes Integration

paleo is battle-tested on Hermes Agent by @mocasus — first user & case study.

# Install via Hermes skills manager
hermes skills install mocasus/paleo

# Or by path
cp -r skills/paleo* ~/.hermes/skills/

Then in your Hermes chat (Telegram, WhatsApp, etc.):

> paleo
🦴 paleo full — terse output, code-first

> build a REST API with FastAPI
[terse, code-first response — no preamble, no filler]

> paleo-auto
🦴 paleo-auto: watching session... enabled paleo + trim-context (23 turns)

> budget 2000
🦴 paleo-budget: 2000 output tokens, hard mode

Hermes tips:

Start with paleo — instant token savings on every reply.
paleo-auto for sessions >15 turns (watches context fill & enables the right skills).
Combine paleo + budget for expensive models via provider routing.
paleo-converse kicks in when your agent loop hits 60%+ context cap.

Custom Skills

paleo is open — wire your own token-saving skills:

skills/<your-name>/SKILL.md with name + description frontmatter.
Add the skills directory to .claude-plugin/plugin.json — the skills field is a path string (e.g. "./skills/"), not an array. Gemini + other agents pick skills up natively; no extra manifest needed.
Bump version badge (this file + footer) + plugin version.
Commit + push.

No repo edit needed — just drop any SKILL.md into your agent's skills dir (e.g. ~/.hermes/skills/<name>/). paleo loads whatever it finds under skills/.

User Stats

Share your numbers, get listed. PR your monthly token savings to this table.

User / Team	Agent	Tokens/month saved	Skills
Add yours →	—	—	—

How to measure:

Use agent 1 week without paleo → note token usage from provider dashboard
Enable paleo (or paleo-auto) 1 week → note new usage
Diff × 4 = estimated monthly savings

Contributing

Contributions are welcome — new token-saving skills, better triggers, or benchmark data.

Open an issue describing the skill or improvement.
Keep SKILL.md files terse (they load into context).
Add name + description frontmatter and register in both plugin manifests.
Bump the version badge and version fields before opening a PR.

License

MIT — see LICENSE.

Paleo

README