crystallize
/crystallize
Use when asked to crystallize a workflow, skill, codebase, or agent runs, or to find where agents repeatedly rediscover methods, apply stable rules, or do unnecessary work. Analysis of opportunities to replace repeated agent effort with reusable procedures; not a general code cleanup or an automatic implementation run.
- Author
- ratacat ↗
- Repository
- ratacats-skills ↗
- Licence
- MIT
- Last repository change
- 22.10.2026
Assessment and security
- Files in this package
- 2
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 22.10.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Use when asked to crystallize a workflow, skill, codebase, or agent runs, or to find where agents repeatedly rediscover methods, apply stable rules, or do unnecessary work. Analysis of opportunities to replace repeated agent effort with reusable procedures; not a general code cleanup or an automatic implementation run.
Original author description
Use when asked to crystallize a workflow, skill, codebase, or agent runs, or to find where agents repeatedly rediscover methods, apply stable rules, or do unnecessary work. Analysis of opportunities to replace repeated agent effort with reusable procedures; not a general code cleanup or an automatic implementation run.
How it works
- Find the repeated work
- Choose the smallest replacement
- Find where the replacement breaks
- Return a short ranked list
- Example
SKILL.md preview
---
name: crystallize
description: Use when asked to crystallize a workflow, skill, codebase, or agent runs, or to find where agents repeatedly rediscover methods, apply stable rules, or do unnecessary work. Analysis of opportunities to replace repeated agent effort with reusable procedures; not a general code cleanup or an automatic implementation run.
metadata:
category: developer tools
blurb: Finds repeated agent work that can be removed or turned into reusable procedures, with evidence, exception conditions, and a small first change.
keywords:
- crystallize
- workflow-distillation
- repeated-reasoning
- agent-workflows
- automation
- cost-reduction
---
# Crystallize
Find where an agent repeatedly figures out something the system could learn to do once. Recommend small replacements that preserve the task's outcome and leave open-ended work to an agent where needed.
Default to analysis and proposals. Implement only when the user's request includes implementation. Use existing code, instructions, records, and run histories; do not require a new tracing system or workflow engine.
## Find the repeated work
Establish the requested outcome and the consequences of a wrong result before proposing shortcuts. Read the relevant procedure, code, or skill, then inspect available runs, including failures and exceptions. Follow an input through the steps that produce the outcome.
Look for repeated reasoning with a stable method, even when each case has a different answer. Identify what stays fixed, what varies, and what the next step actually needs. Cite the specific instructions, code, or runs behind each finding.
Code and prompts reveal candidates. Runs establish recurrence and expose variation. Without run evidence, label the opportunity a hypothesis and identify the smallest observation that would settle it. Never invent frequencies, costs, or savings.
## Choose the smallest replacement
Consider these changes, starting with work that can disappear:
- Remove a step when its result cannot affect any required outcome. Include required evidence and reporting in that outcome; an unchanged verdict alone may not justify skipping research.
- Reorder work or stop early when the remaining results cannot change the outcome. State why that condition holds and what would invalidate it.
- Retain a discovered method, address, query, or retrieval route. State when it must be refreshed or rediscovered. Reusing a method does not make yesterday's facts current.
- Turn explicit rules, parsing, calculations, polling, and mechanical actions into ordinary code or an existing tool.
- Isolate a narrow judgment with clear inputs and possible answers. Carry the relevant criteria, exceptions, and blockers into that judgment. Keep uncertain cases on a path that can resolve them.
- Keep an agent for investigation or decisions whose scope remains open. A short reusable instruction can be sufficient when code would be brittle.
Do not prescribe a model vendor, DSL, or framework. Use the project's existing tools where they fit. A model's stated confidence is not evidence of calibration; any routing threshold needs support from representative outcomes.
## Find where the replacement breaks
For each serious candidate, look for an actual exception or construct a clearly labeled counterexample. Repeated successful answers may share an accidental condition.
Describe the inputs the replacement handles and the observable condition that returns work to an agent or person. Identify silent failures as well as explicit errors. If the system cannot recognize the exception reliably, narrow the proposed replacement or retain the current investigation.
Do not assume the current agent's output is correct. Compare against the task's requirements and reviewed outcomes where available. Propose a bounded comparison using ordinary cases and relevant exceptions. Use cases outside those that suggested the shortcut when available. Preserve any asymmetry in the cost of mistakes.
Retained evidence can support replay only for decisions it still contains the necessary facts to answer. A new question, policy requirement, or freshness requirement may need new research.
## Return a short ranked list
Rank by recurrence, avoidable effort, consequences of mistakes, and maintenance burden. Distinguish observed savings from estimates, and account for fallback work and upkeep. Prefer a few supported opportunities; no worthwhile candidate is a valid result.
For each recommendation, give:
- The repeated work and its evidence.
- The stable part and the smallest replacement.
- The exception condition and where that case goes.
- The likely benefit, with the basis for any estimate.
- The smallest comparison or observation needed before adopting it.
Recommend one starting point. Keep speculative candidates visibly separate from findings supported by runs. Return the analysis in the conversation unless the user requests another artifact.
## Example
An agent repeatedly navigates an organization's website to find the same calendar. Save the calendar address and retrieval method so later runs can start there. Return to discovery if retrieval fails or the page no longer contains the expected listings. Check freshness separately; a successful fetch may return an abandoned calendar. Compare later runs against independent discovery before treating the shortcut as dependable.
Files in this package (2)
- README.md 1 KB
- SKILL.md 5 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · ratacat/ratacats-skills | 100 | MIT |
| kirodotdev/kirocrew | 100 | Apache-2.0 |
| eric-tramel/moraine | 100 | Apache-2.0 |
Database snapshot: 13.11.2026 · package version 6b5832453131 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100agent-skill-author
Create, improve, audit, and package reusable Agent Skills that follow the open SKILL.md format. Use when turning a workflow or domain procedure into a portable skill for compatible AI agents. - 100create-skill-claude
Use this skill when the user asks to "create a skill", "add a skill", "make a new skill", "build a skill", or wants to automate a repeated workflow into a reusable prompt. Guides creating a properly structured Claude Code skill. - 86skill-authoring
Create or update reusable Crumb Agent Skills when the user asks for a custom skill, workflow, capability pack, or SKILL.md. Do not activate for ordinary implementation work that merely uses an existing skill. - 100run-history-skill-builder
Turn a completed task, browser flow, artifact pipeline, failure-recovery trace, or repeatedly refined workflow into a new reusable skill package or a reviewed skill-design plan. Use when the user asks to make a new skill from real run history, extract a reusable workflow from… - 100session-to-skill
Use when the user wants to turn one or more Entire-tracked sessions, checkpoints, or repeated agent workflows into a reusable agent skill. - 95workflow-skill-creator
Distills a completed user workflow or interaction into a reusable agent skill. Use when the user asks to turn their workflow, interaction, or multi-step process into a skill, or when they say "make this a skill", "create a skill from what we just did", "package this workflow" or…