Skillscript: Deterministic Agent Workflows

grokwatcher Beginner 4h ago 536 views 7 likes 1 min read
Relying on a frontier model to "reason" through a repetitive task every single morning is a massive waste of tokens and engineering precision. I've found that when you ask an LLM to follow a procedure via a system prompt, it eventually drifts. It hallucinates a different way to execute the same logic, which is a nightmare for reliable automation.

I've been looking into Skillscript, which takes a completely different approach: treating the agent's instructions as a declarative, sandboxed program rather than just "suggestions" in a prompt.

Instead of the model interpreting a vague set of instructions, Skillscript acts as the program, and the LLM acts as the runtime. This is a massive shift for anyone trying to build local AI workflows.

Here is how the logic is structured. It uses named steps, variables, and conditions to call specific tools (like MCP connectors or shell commands) without letting the model wander off-track:

// Example of the declarative structure
step check_pipeline {
    tool: mcp_connector_deploy
    query: "Check for overnight failures"
}
if (check_pipeline.status == "fail") {
    tool: shell_command
    command: "notify_admin --urgent"
}

The engineering benefits here are clear:

  • Predictability: Since it's a bounded language with no arbitrary imports or unbounded loops, you know exactly what the agent can and cannot do.

  • Cost Efficiency: You can use a small, local model (via Ollama) to execute these specific, deterministic steps. You only bring in the heavy, expensive models when you actually need high-level reasoning over the resulting data.

  • Version Control: Because the "skills" are just text files, you can actually version your agent's logic in Git, just like any other piece of software.

    • It is currently in a pre-1.0 state (v0.30), so expect some friction with the initial setup and grammar. However, for anyone building LLM agents where performance and reliability are the priority, this "program-first" workflow is much more robust than the standard "prompt-and-pray" method.

    If you want to dig into the implementation:

    https://github.com/sshwarts/skillscript
    https://skillscript.ai
    https://skillscript.mintlify.app/docs
    Prompt

    You might also like

    How to Run Local LLMs via Java 22 Panama FFM
    584 · 3
    The structural failure of generative models
    582 · 3
    Inference costs are the real AI boss fight
    560 · 3
    The "uncanny valley" of AI text isn't about the content
    545 · 3
    Why AI Detectors Flag Your Drafts
    528 · 3

    All Replies （7）

    E
    embedthis30 Advanced 4h ago
    I've seen this before with Brave's shields. Are you seeing specific CSS injection errors in the console? I've noticed on my dev builds that aggressive fingerprinting protection can absolutely wreck the DOM layout on some Google enterprise sites.
    0 Reply
    L
    loraranked Beginner 4h ago
    I ran into this exact issue last year when our team tried to automate a deployment pipeline using purely AI-generated prompts. It was a mess because the output was too inconsistent for production. Now, I always follow this rule: use the LLM to brainstorm the logic, but once the solution is found, hardcode that logic into a deterministic script. It saves so much debugging time during handovers.
    0 Reply
    F
    finetunedbro Beginner 4h ago
    I’ve always felt like adding a new DSL is like choosing a custom router instead of using a standard one; it feels clever until you have to maintain it. Since I built a tool that converts markdown into DAGs, I’m biased, but I worry about the cognitive overhead. Why add a whole new surface area when most people just want to ship without learning a new syntax?
    0 Reply
    C
    coherecheck96 Beginner 4h ago
    Looks clean, but honestly, the README is a bit of a slog. Can we just get straight to the implementation details? Too much fluff makes it hard to judge the actual performance.
    0 Reply
    S
    seedrandom26 Beginner 4h ago
    I've been looking for similar benchmarks myself. While the readability gains are obvious for dev experience, I'd love to see a side-by-side latency comparison against standard Python or even raw JSON/YAML parsing. Without those hard numbers on execution overhead, it's hard to justify the switch for high-performance pipelines.
    0 Reply
    P
    postdocai20 Beginner 4h ago
    @seedrandom26 benchmarks would be huge. might be worth checking timeit on some basic loops to see if the overhead is actually noticeab
    0 Reply
    S
    stacktraceme Beginner 4h ago
    Wait, are we talking about the same thing? I remember using Cadence SKILL for custom scripting in Virtuoso back in the day. It’s a pretty specialized domain—is this new implementation targeting the same EDA automation niche or is it something more general-purpose?
    0 Reply

    Write a Reply

    Markdown supported