Why Skillscript?
Frontier reasoning is the scarce, expensive thing, and agents can squander it
Skillscript came about from building out my personal NanoClaw based agent, Perry, our workflow, and facing the challenges we had to solve along the way. While our shared memory system helps, it only goes so far.
The two problems that emerge right away are:
- Re-doing expensive work that should be captured once (memory upkeep, re-reasoning, rewriting throwaway tools).
- Not having the right thing ready at the moment it's needed (guessing relevance on a subject change, not knowing how to handle an event).
Both are the same waste: a frontier model that spends its reasoning on rote re-derivation or relevance guessing instead of focusing on what’s genuinely novel.
That’s why we built Skillscript. The bet is that you can capture work once as a compiled, auditable, reusable skill and stop paying a frontier model to re-reason it every time.
Here’s how we use Skillscript:
Memory Maintenance
Every morning at 3 am, a Skillscript cron fires. The script it executes calls a local LLM that classifies memories and writes a digest of stale threads, blocked projects, and inconsistent states for Perry to review on his morning wake-up. If there are any serious issues with the memory store, Perry is awoken with the diagnostic.
Morning Briefing
At 8am, a cron job fires, causing Perry to execute his morning-brief skillscript. That script does the following:
- Executes another skillscript that retrieves the daily weather
- Checks Perry’s mailbox for new items
- Retrieves any actionable items from the memory maintenance run
- Executes a skillscript that uses a local LLM to check for any mentions of Skillscript on the web
- Checks the Skillscript GitHub for any submitted issues
Project Brief
Any time Perry is asked for a project brief, he fires off a project-brief skillscript that:
- Retrieves an anchor summary from the memory store
- Checks the issue tracker (YouTrack) for any new issues
Get Attention
If Perry (or any other agent) needs to get my attention, they can use the OSX ‘say’ shell command. Since Perry is in a container, he can’t call ‘say’ so he merely executes a script, passing the words to be spoken.
Event-Driven Recovery
- A health-check webhook posts to the Skillscript runtime (no cron, no human).
- It restarts the failing service, re-checks, and if it fails, wakes Perry with a diagnostic.
Triage-Advisor
A ticket comes in. Instead of re-reasoning triage every time, Perry asks a triage-advisor skill — written and approved once — "what priority is this, and what's the playbook?"
- The script branches deterministically; or it first asks a local LLM to classify the text, then branches.
- The flip that makes it interesting: the agent supplies the input and asks the script what to do. Judgment captured, not re-derived.
One Command Deploy
This website. A single skillscript re-deploys the website vs a multi-step curl-over-SFTP dance.