auto-route

Recommend or switch to a token-saving model tier (fast / default / premium / local / gateway) based on task hardness. Triggers: "auto route", "save tokens", "use cheap model", "route this task", "which model should we use".

Jump to install

Repository: hoodini/logan-cli
Last source activity: July 15, 2026 at 23:48
Detected SKILL.md language: English
Stars: 9
Forks: 1

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

A direct command skips the review prompt. Inspect the source before running it.

npx skills add https://github.com/hoodini/logan-cli --skill auto-route

The command stays on one line. Scroll horizontally to inspect it before copying.

Prefer a local copy? Download the files currently available to SkillsMP.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

SKILL.md

Source instructions · Read-only preview

Back to search Back to installation

self-improve

Visibility into Logan self-healing and self-improving. Show what was analyzed, what improved, how, and why. Journal durable lessons. Trigger: /improve, /heal, /reflections, "what did you learn", "why did you decide that".

2026-07-19

yuv-design-dna

Yuval's design DNA - MANDATORY for any web/frontend deliverable: landing page, website, app UI, dashboard, React, HTML, CSS, portfolio, promo page. Signature moves (video-scrub heroes, parallax frame-scrub, HyperFrames motion), the YUV design system (palettes, Anton/Inter + Rubik/Assistant, radii 0/999), hard accessibility + mobile-first gates, and the anti-AI-look checklist. Trigger: landing page, website, frontend, UI, dashboard, React, HTML, CSS, design, page, app. Hebrew: אתר, דף נחיתה, עיצוב, דאשבורד, אפליקציה.

2026-07-19

yuvai-thinking

Yuval Avidani's learning-and-teaching method. Use WHENEVER Yuval wants to learn, understand, or be taught anything, AND when building educational/explanatory content (tutorial, course, interactive lesson, deep-dive explainer, walkthrough). Governs HOW to explain: every crumb, the why behind each thing, intuition before formula, ZERO forward-references, zero assumed prior knowledge, real interactive examples, full under-the-hood process (never 'just use a library'). TRIGGERS — English: explain, teach me, what is, how does X work, help me understand, deep dive, walk me through, from scratch, I don't understand, you skipped/forgot/didn't explain. Hebrew: תסביר, תלמד אותי, מה זה, איך עובד, תעשה לי סדר, לעומק, מההתחלה, מאפס, פרק לי, תבנה לי שיעור/קורס/הסבר, לא הבנתי, דילגת, שכחת, לא הסברת, החסרת, הנחת. Also trigger when Yuval is frustrated an explanation was incomplete, jumped ahead, or assumed knowledge. Not domain-specific — apply to anything Yuval learns.

2026-07-17

View all from this repository

name	auto-route
description	Recommend or switch to a token-saving model tier (fast / default / premium / local / gateway) based on task hardness. Triggers: "auto route", "save tokens", "use cheap model", "route this task", "which model should we use".
model	tier-fast
effort	low

Auto-route (Logan)

Pick the cheapest model that can still succeed. Tiers are defined in examples/config/auto-routing.toml (tier-fast, tier-default, tier-premium, tier-local, tier-gateway).

Heuristics

Tier	When
fast	Q&A, explain, search-only, 1-file tiny edit, classify
default	Normal implement, tests, refactors under ~15 files
premium	Architecture, multi-package, nasty bugs, security
local	Offline / privacy / bulk offline classification
gateway	When LiteLLM is the org standard path

Steps

Estimate task size: files touched, ambiguity, risk of prod break.
Recommend a tier with one-sentence why.
If the user agrees (or said "auto"), switch with /model <tier-id> when available in this session, or tell them to re-run with -m <tier-id>.
For multi-step work: start fast for exploration, escalate to default/premium only when blocked.
After finish, note estimated savings vs always-premium (qualitative is OK).

Rules

Never route security-sensitive production changes to an untrusted local model without user OK.
Prefer escalating once over thrashing models every turn.
Combine with /context awareness - long contexts are expensive; compact or flush before premium if history is bloated.

auto-route

Source facts

Install options

Review the source files

More from this repository

More from this repository

Auto-route (Logan)

Heuristics

Steps

Rules

Auto-route (Logan)

Heuristics

Steps

Rules