create-ex

Overview of create-ex skill

What create-ex does

The create-ex skill is a guided workflow for turning relationship context plus chat history into a reusable digital persona skill. It is built for users who want more than a one-off imitation prompt: the real job is to extract tone, conflict patterns, care signals, and reply habits from WeChat, iMessage, or pasted transcripts, then write those patterns into files a coding agent can reuse.

Who should use this create-ex skill

Best fit: users comfortable sharing structured personal context and message history, especially if they want a repeatable persona artifact rather than a single simulated chat. It is particularly relevant for create-ex for Prompt Writing, because the repo contains separate prompts for intake, chat analysis, persona analysis, persona building, and correction handling instead of hiding everything in one monolithic instruction.

What makes it different

Compared with an ordinary prompt, create-ex has a staged pipeline: intake, data import, automatic analysis, preview, then file output. It also includes concrete prompts such as prompts/chat_analyzer.md, prompts/persona_analyzer.md, and prompts/persona_builder.md, plus example output under exes/example_liuzhimin/. That structure reduces guesswork if you want a persona you can inspect, edit, and improve later.

Key adoption checks first

Before a create-ex install, check three things: whether you are okay processing sensitive chat logs, whether your environment can run Python 3.9+ tooling, and whether your main source is supported. The repo clearly supports WeChat and iMessage workflows, but WeChat extraction is legally and operationally unstable per the README, so some users will rely on pasted text, screenshots, or custom export work instead of full automation.

How to Use create-ex skill

create-ex install and setup path

There is no single canonical package install command in SKILL.md, so treat this as a repo-based skill. Start by reading SKILL.md for the trigger flow, then README_EN.md, then docs/PRD.md. If you want the bundled tooling, you will also need Python 3.9+ and should inspect tools/wechat_decryptor.py, tools/wechat_parser.py, and tools/skill_writer.py before trusting the automation. The README shows a clone-based install path for OpenClaw:
git clone https://github.com/titanwings/ex-skill ~/.openclaw/workspace/skills/create-ex

What input create-ex needs

Good create-ex usage depends on two input layers:

1. short human context: name/codename, relationship stage, age range, traits, MBTI, attachment style, astrology if known;
2. evidence: chat logs, screenshots, pasted transcripts, or extracted messages.

Stronger inputs produce better persona rules. For example, “She is caring but avoidant” is weak. Better: “Female, 25, dated 8 months, breakup; usually replies fast when calm, goes silent during conflict, says ‘I’m not that kind of person’ when defensive, often uses ellipses and ‘haha’ to soften criticism.”

How to prompt create-ex well

To invoke create-ex effectively, give a rough goal plus operating constraints up front. A strong request looks like:

• Goal: “Use /create-ex to build a reusable persona skill, not just sample dialogue.”
• Source: “Primary evidence is pasted WeChat text from the last 6 months.”
• Priority: “Preserve texting rhythm, conflict behavior, and affection style.”
• Constraint: “Mark low confidence if evidence is thin; quote original lines when possible.”

This maps well to the repo’s internal logic: manual labels override transcript inference, low-volume samples should be flagged, and claims should be grounded in quoted messages.

Best workflow and files to read first

Use this order:

1. SKILL.md — operational steps and trigger behavior
2. prompts/intake.md — what metadata the skill expects
3. prompts/chat_analyzer.md — what the transcript parser should extract
4. prompts/persona_builder.md and prompts/correction_handler.md — how the final persona evolves
5. exes/example_liuzhimin/ — what finished output looks like

Practical tip: do not start with screenshots if you already have text. The analysis prompts assume message content can be classified into long messages, conflict messages, affectionate messages, and daily chat. Clean text gives far better downstream structure than OCR-heavy screenshots.

create-ex skill FAQ

Is create-ex better than a normal imitation prompt?

Usually yes, if you need continuity. A normal prompt can mimic surface tone for one session, but create-ex is designed to generate a stored persona with analysis layers and correction handling. That matters when you want to refine behavior over time instead of rewriting the same instructions every session.

Is this create-ex skill beginner friendly?

Moderately. The high-level flow is simple, but the repo mixes prompt assets with Python tools and assumes some comfort reading files and adapting workflows. Beginners can still use create-ex by focusing on manual text input and the prompt files, skipping automated extraction if local setup feels risky or unclear.

When should I not use create-ex?

Skip it if you only need a playful one-off conversation, if you cannot safely process private chats, or if you expect factual psychological diagnosis from sparse messages. It is also a weak fit if your source material is tiny; the repo itself warns that fewer than 200 messages lowers confidence.

How does it fit prompt-writing workflows?

create-ex for Prompt Writing is a strong use case because the repository exposes reusable prompt components. You can borrow the intake, analysis, and correction patterns for adjacent persona-building projects, even if your final output is not specifically an “ex” simulation.

How to Improve create-ex skill

Give create-ex higher-signal evidence

The biggest quality lever is evidence quality, not more adjectives. Feed representative messages across calm, affectionate, conflicted, and distant periods. Include timestamps or rough phases if possible. If you know certain traits are definitely true, state them explicitly, because the skill prioritizes manual labels over uncertain inference.

Reduce common create-ex failure modes

Most weak outputs come from one of four issues: too little data, over-romanticized user descriptions, transcript fragments without speaker clarity, or no correction pass after preview. If the first draft feels generic, that is often because the source only reveals mood, not repeated language habits or conflict sequences.

Iterate after the first output

The repo includes prompts/correction_handler.md for a reason: use it. Instead of saying “this feels off,” give scenario-specific fixes such as: “When angry, they do not explain; they stop replying for hours and return with a practical message.” Concrete corrections can be written into the persona and are more valuable than broad style complaints.

Adapt the workflow, not just the theme

To improve create-ex, treat it as a persona-construction framework. If WeChat automation is blocked, use the intake prompt plus manually curated transcript text. If the final persona overfits cute phrases, rebalance by adding more conflict and routine chat. If you are using create-ex guide patterns for another domain, keep the staged design: intake → evidence import → analysis → preview → correction → saved artifact.