create-crush

Overview of create-crush skill

What the create-crush skill actually does

The create-crush skill is a Claude Code skill that turns raw personal material into a reusable character-style skill for one person: chat exports, photos, social posts, and your own descriptions become a structured Relationship Memory plus Persona. The real job-to-be-done is not “make a chatbot,” but “preserve how this person feels, sounds, and relates to me” in a form you can refine over time.

Who should install create-crush

create-crush for Skill Authoring fits users who want a guided pipeline instead of improvising one long prompt. It is best for people already using Claude Code, comfortable placing files in .claude/skills/, and willing to provide source material. If you only want a one-off roleplay reply, a normal prompt is faster; if you want a maintainable, updateable persona with memory correction, the create-crush skill is more useful.

Why it stands out from a generic prompt

Unlike a plain “act like this person” prompt, create-crush includes intake prompts, parsers for WeChat/QQ/social content, optional photo analysis, generated files under ./crushes/{slug}/, and a correction workflow for “that’s not how they talk.” That structure reduces guesswork when building or evolving the persona.

Important fit and boundaries before install

This skill is explicitly framed for personal memory and emotional analysis, not harassment, stalking, or privacy invasion. Adoption blockers are practical: it expects Claude Code, optional Python 3.9+ for helper scripts, and source material in formats the included tools can parse. If you cannot or should not process private data locally, do not install create-crush.

How to Use create-crush skill

Install create-crush in the right Claude Code location

For create-crush install, clone the repo as a skill folder named create-crush:

# project-local
mkdir -p .claude/skills
git clone https://github.com/xiaoheizi8/crush-skills .claude/skills/create-crush

# or global
git clone https://github.com/xiaoheizi8/crush-skills ~/.claude/skills/create-crush

# optional helpers
pip3 install -r .claude/skills/create-crush/requirements.txt

The main environment assumption is Claude Code with Read, Write, Edit, and Bash available. Optional Python dependencies mainly improve parsing, not core prompting.

Start with the files that matter most

For a fast create-crush guide, read these in order:

1. SKILL.md — trigger phrases, tool rules, output location, update mode.
2. README_EN.md — install path and user-facing commands.
3. prompts/intake.md — what the skill will ask you first.
4. prompts/memory_analyzer.md and prompts/persona_builder.md — what gets extracted.
5. prompts/correction_handler.md — how post-generation fixes are applied.

If you plan to import data, also inspect tools/wechat_parser.py, tools/qq_parser.py, tools/social_parser.py, and tools/photo_analyzer.py so you know what formats are realistically supported.

Trigger create-crush and give stronger inputs

Typical create-crush usage starts with /create-crush or natural language like “Help me create a crush skill.” The skill then needs:

• a codename or nickname
• a short relationship summary
• a personality description
• optional files: chat logs, screenshots, photos, social text

Weak input: “She is nice and cute.”
Better input: “Codename: Xiaomei. We met 3 months ago at work in Shanghai. We mostly chat after 10pm. She replies briefly in text but is warmer in person. She rarely initiates, uses few emojis, and often ends with ‘hhh’.”

That stronger input improves output because the skill can separate facts, interaction patterns, and speech habits instead of inventing them.

Use the generated skill as a workflow, not a one-shot

After creation, the repo supports a lifecycle: create, test, correct, and evolve. Useful commands mentioned in the docs include /list-crushes, /{slug}, /{slug}-memory, /{slug}-persona, and update flows such as /update-crush {slug}. Best practice:

1. create from minimal facts plus a few representative materials
2. test dialogue
3. correct factual errors and tone errors separately
4. add more logs later rather than overloading the first run

This staged workflow is the main reason to choose create-crush over a single prompt.

create-crush skill FAQ

Is create-crush better than ordinary prompting?

For repeated use, yes. A normal prompt is fine for a quick imitation, but create-crush gives you structured memory, persona layers, import helpers, and explicit correction handling. That matters when you want consistency across sessions or want to keep improving the result.

Is the create-crush skill beginner-friendly?

Moderately. The interaction flow is guided, but installation is not no-code: you need Claude Code skill folders and may want Python for parsers. If you are new to Skill Authoring, this is usable, but not the simplest first skill because it mixes prompts, local files, and privacy-sensitive source material.

What are the main limits of create-crush usage?

The quality depends heavily on your materials. Sparse, idealized, or contradictory inputs will produce flatter personas. The repo also appears optimized for Claude Code-style tool use, so it is not a generic web app or hosted service. It can preserve patterns from evidence; it cannot reliably infer hidden feelings or fill major gaps safely.

When should I not use create-crush?

Skip create-crush if you need a public-facing character bot, a compliance-safe enterprise workflow, or anything involving consent problems or invasive data collection. It is also a poor fit if your goal is broad fictional persona generation rather than a memory-grounded, evidence-driven representation of one real person.

How to Improve create-crush skill

Give evidence in layers, not a dump

To improve create-crush, provide inputs that map cleanly to the repo’s structure:

• facts: how you met, timeline, current status
• behavior: who initiates, reply speed, favorite topics
• language: recurring phrases, emoji habits, message length
• moments: a few memorable interactions with context

This is better than uploading everything with no framing. The analyzers work best when the model can separate “what happened” from “how they sound.”

Fix the most common failure modes early

The biggest errors are usually:

• overly romantic or overly confident tone
• generic speech patterns
• confusing your projection with observed behavior
• mixing one-off events into stable traits

Use the correction flow for precise fixes: “She would not send long emotional paragraphs,” or “We did not meet alone; update the relationship timeline.” The included correction handler is one of the highest-value parts of the create-crush skill.

Iterate with representative samples

Do not judge the first output only on whether it feels emotionally right. Test specific cases:

• small talk
• late-night reply style
• awkward distance
• memory recall of shared moments

Then add more representative data from the modes where it fails. Five realistic exchanges are usually more useful than fifty bland ones.

Improve create-crush for Skill Authoring reuse

If you want to adapt create-crush for Skill Authoring, study how SKILL.md delegates specialized work to prompt files and Python tools. The reusable pattern is: intake → parse sources → build memory/persona → write a target skill → support corrections/versioning. That architecture is valuable even if your own skill is about mentors, fictional characters, or customer personas instead of a crush.