kasperjunge

refactor-for-determinism

Design or refactor skills by separating deterministic and non-deterministic steps. Use when creating or improving skills, especially to move repeatable workflows into scripts/ and update SKILL.md to call them.

kasperjunge 450 35 Updated 7mo ago
GitHub

Install

npx skillscat add kasperjunge/agent-resources/refactor-for-determinism

Install via the SkillsCat registry.

About this skill

This skill helps organize AI agent workflows by isolating repeatable, mechanical tasks into scripts, leaving judgment-based steps in the main documentation. It improves reliability by automating deterministic actions like version checks or diff collection, reducing errors from manual processes. Use it when building or refining skills to enhance consistency and maintainability.

SKILL.md

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