What counts as a deterministic step?

A repeatable, mechanical task that can be scripted, validated, and that should fail clearly if inputs or state are invalid (e.g., running quality checks, verifying a clean git state, updating version strings, promoting CHANGELOG sections, collecting diff context).

Where should scripts live?

In the skill's scripts/ directory. Each script should be self-contained, provide clear error messages, and exit non-zero on failure.

How do you handle steps that involve judgment?

Keep all judgment or creative decisions in SKILL.md prose; call out exact boundaries with a Deterministic Steps vs Non-Deterministic Steps structure and reference the scripts only for scripted parts.

refactor-for-determinism

npx machina-cli add skill kasperjunge/agent-resources/refactor-for-determinism --openclaw

Files (1)

SKILL.md

2.6 KB

Refactor for Determinism

Build reliable skills by separating deterministic steps from judgment-based steps.

Core Principle

Deterministic steps belong in scripts. Use SKILL.md to orchestrate the workflow and reserve judgment for the non-deterministic parts.

Workflow

1. Identify Deterministic vs Non-Deterministic Work

For each step in the skill:

Deterministic: repeatable, mechanical, or validation-heavy steps → script candidates
Non-deterministic: judgment, interpretation, creative choices → keep in SKILL.md

Examples of deterministic steps:

Running quality checks
Verifying clean git state
Updating version strings
Promoting CHANGELOG sections
Collecting diff context for review

Examples of non-deterministic steps:

Writing changelog content
Selecting a solution approach
Code review judgments
Deciding release timing

2. Design Scripts for Deterministic Steps

For each deterministic step:

Create a script in scripts/ within the skill directory
Make it self-contained with clear error messages
Validate inputs and exit non-zero on failure
Prefer small, single-purpose scripts

3. Update SKILL.md to Use Scripts

Replace manual command lists with script calls
Reference scripts using relative paths: scripts/...
Keep judgment steps explicit in prose

4. Document Boundaries

Make the line between scripted and non-scripted steps obvious:

Use section headers like "Deterministic Steps" and "Judgment Steps"
Call out where human/agent judgment is required

Output Format

## Determinism Audit

### Deterministic Steps (script candidates)
- [Step] → [script name]

### Non-Deterministic Steps (keep in SKILL.md)
- [Step] → [why it needs judgment]

### Script Plan
- scripts/[name] - [purpose, inputs, outputs]

### SKILL.md Updates
- [Where to call each script]

Common Mistakes

Mistake	Fix
Scripting judgment	Keep decision-making in SKILL.md
One giant script	Split into small, focused scripts
Silent failures	Print clear errors and exit non-zero
Hardcoded paths	Use repo-relative paths
Forgetting SKILL.md updates	Always wire scripts into instructions

What NOT to Do

Do NOT hide decisions inside scripts
Do NOT make scripts that require manual editing
Do NOT mix multiple responsibilities into one script
Do NOT add extra documentation files beyond SKILL.md

Source

git clone https://github.com/kasperjunge/agent-resources/blob/main/skills/skill-development/refactor-for-determinism/SKILL.md

View on GitHub

Overview

Refactor for determinism helps design or refactor skills by clearly separating deterministic, repeatable steps from non-deterministic judgment. It advocates moving repeatable workflows into scripts/ and uses SKILL.md to orchestrate the workflow and document where human judgment is required.

How This Skill Works

Identify which steps are deterministic (repeatable, mechanical, or validation-heavy) and which require judgment. Create small, self-contained scripts under scripts/ for the deterministic steps, with clear error messages and input validation that exit non-zero on failure. Update SKILL.md to call these scripts using repo-relative paths (e.g., scripts/...), while keeping judgment-heavy decisions in prose to maintain clarity.

When to Use It

Starting a new skill that includes repeatable checks, builds, or validations.
Refactoring an existing skill to isolate deterministic steps into scripts for reliability.
Preparing a release where version bumps, changelog handling, and diff collection are automated.
Auditing a skill to clarify where scripted steps end and human judgment begins.
Standardizing cross-skill workflows by centralizing deterministic tasks into reusable scripts.

Quick Start

Step 1: Identify deterministic vs non-deterministic steps in your skill's workflow.
Step 2: Create small, self-contained scripts under scripts/ for deterministic steps and wire them into the workflow.
Step 3: Update SKILL.md to call the scripts with repo-relative paths and keep judgment steps in prose.

Best Practices

Clearly label steps as Deterministic (scripted) or Non-Deterministic (judgment).
Place small, single-purpose scripts under scripts/ and guard each with clear error messages.
Validate inputs in scripts and exit non-zero on failure to prevent silent errors.
Reference scripts in SKILL.md using relative paths like scripts/...
Keep judgment steps in prose and avoid hiding decisions in scripts.

Example Use Cases

An automated skill validates git state, runs tests, updates version strings, and captures diff context via scripts.
A CHANGELOG promotion step is scripted while the rationale for entries remains described in SKILL.md.
A deterministic packaging or release-prep task is automated with a small script, producing consistent outputs.
Diff-context collection for code reviews is automated by a script rather than written manually.
Scripts verify inputs and environment constraints to ensure consistent behavior across runs.

Frequently Asked Questions

Add this skill to your agents