Skip to content

Ghost 1.6: skills that cut tokens by 90%

Ghost 1.6 cuts agent tokens by a median of 90% when your agent replays a skill instead of working through the app again. That’s about 10x fewer, measured across 99 skills, and the tasks finished a bit faster too.

The rebuilt skill system ships 117 skills for 113 Android apps, up from the 3 skills in the previous hub. Your agent learned an app once. Now it doesn’t have to relearn the same screens on every run.

Ghost in the Droid gives your agent a phone to operate. It reads the screen, taps, swipes, and types. GHOST is the agent haunting that phone, and this release gives it a library of routes to call when the way through an app is already known.

Hard skills replay the route, soft skills keep the guide

Section titled “Hard skills replay the route, soft skills keep the guide”

A hard skill is an executable YAML action pinned to an OS version, device, and app version. It replays without an LLM, so your agent calls the action instead of reasoning through each tap again.

The pinning matters. A workflow recorded against one app layout can break when the layout changes, and a successful run on the author’s phone doesn’t show that the same action works on yours.

Every skill also ships a soft skill: a plain-language SOFT.md guide to completing the task. When no hard skill matches your setup, your agent reads that guide, operates the app itself, and can then create a fitting hard skill for your phone.

In short, the hard skill executes a known route, and the soft skill explains the task when your agent has to find its way.

A median of 90%. Every one of the 99 skills with a before/after recording ran the same task twice on a real phone, with the same prompt and the same model (Haiku 5.5): once without the skill, once with it. The median token reduction was 90%, a 10.2x reduction factor (95% interval 6.7x to 11.8x). In 4 of the 99, the skill used more tokens than driving by hand. I’d rather you hear that from us.

The task also finished faster, a median 1.24x.

A separate controlled run on 39 actions, one phone, repeated with and without the skill, points the same way: a median 85% fewer tokens.

With vs without a skill (without = 100%) Tokens without 100% with 10% (90% fewer) Time without 100% with 81% (19% less time) Tool calls without 100% with 29% (71% fewer) Median over 99 skills, Haiku 5.5, real phones, same prompt with and without the skill.

The Calculator demo shows it on one task: 92% fewer tokens with the skill.

A soft skill isn’t a shortcut. It’s the fallback, and the starting point for a new hard skill.

When no hard skill matches your phone, Android version or app version, your agent reads the skill’s SOFT.md: what the task is, which screens to expect, where the traps are. It drives the app by hand and still gets the job done. That run costs tokens like any agent run. Then your agent can write down the route it found as a new hard skill for your setup, following the skill creation guide. From then on, that action replays without an LLM.

Soft skill to find the route once, hard skill to replay it. See Hard vs Soft skills in the docs.

Maybe, and the honest answer is a matrix. A skill is pinned to the setups it was recorded on, so the Hub now publishes every skill against every phone and Android version it was run on, and how often it worked there, failures included.

SetupSkills testedRuns passed
Galaxy A03s, Android 13634,714 / 4,893
Galaxy A15 5G, Android 16383,648 / 3,656
Emulator, Android 1322167 / 194
Emulator, Android 1420239 / 240

Most skills are tested on one setup so far, and 20 run on two or more. A blank cell means not tested yet, not broken. Run a skill on your own phone and your setup gets added. See the full compatibility matrix.

The Skill Hub puts installation and evidence in one place. You can search, install, and update skills from the CLI or through MCP, and before you pick an action you can look at what supports it.

There are demo clips for 99 skills, plus before/after videos for 99 skills. A demo shows the action being performed. A comparison shows the same prompt and phone with and without the skill. Watch what it does, then read the evidence behind it.

Each skill has a score from 0 to 100. Evidence and documentation earn points. Downloads and likes don’t.

Some steps need a person. A skill can pause, ask you to handle a checkpoint such as a captcha, and resume afterward.

So human intervention is part of the workflow. Your agent hands control back when it needs to, then carries on once you’ve dealt with the checkpoint. We tried it on four captcha demo pages (reCAPTCHA, hCaptcha, BotDetect and MTCaptcha): each time a fresh agent filled the form, called for a human, and finished the job once the captcha was solved. Watch the four demos and read how to run an AI agent with a human in the loop on Android.

Upgrade an existing installation and install the Calculator skill:

Terminal window
pipx upgrade ghost-in-the-droid
ghost skill install ghost/calculator

For an existing legacy skill, use the migration command:

Terminal window
ghost skill migrate <path>

Browse the Hub, check the compatibility information, and look at the recorded runs. Start with an action you can verify yourself, and watch where your setup differs from the recorded one, especially the OS, device, and app version.

What we need most is independent evidence: people running skills on their own phones and telling us what happens.

Skill mining and automatic adaptation are planned work. They aren’t shipped features of this release.

For now, Ghost knows more routes through your apps, and you get a record you can inspect before sending it through the screen.

Join the Discord to share the skills and things you build with Ghost, report failures, and get support. Bring your phone. Help teach the ghost.