jinhanbuilds/bushiershi

bushiershi is an Agent Skill that counteracts the model’s poor expression habits.

It only takes effect when explicitly called $bushiershi.

Where does it come from?

I’ve been using GPT for over three years. Over the past few months, I’ve compiled the repeatedly corrected expression issues in real tasks into this Skill.

Read about the origin of ‘not really.’

The Underlying Approach of Human Speech

• Human Speech Comes from the Content Itself: Specific objects, real actions, clear relationships, and well-founded judgments are more important than imitating spoken language, short sentences, or a certain “human-like” tone.
• AI flavor is often a writing action: The problem isn’t just a few words, but also how the model organizes conclusions, creates a sense of completeness, adds meaning, and ends a paragraph.
• Smooth Editing Does Not Come at the Expense of Distortion: Shorter, more direct text, if the facts, degree, attribution, or uncertainty are altered, are simply more confident misspellings.
• Only control the final draft: This skill handles the text readers actually see and does not handle the entire Agent’s thinking, collaboration, or task advancement.

Quick to use

One-click installation

Run in the Git project directory where you want to use the skill:

This command installs the Skill into the current project’s .agents/skills/bushiershi.

Manual installation

When not using the Skills CLI, download this repository and run it in the root directory:

If your client uses another Skills directory, replace the destination path with the corresponding location.

Invoking

Explicitly call $bushiershi] in the request:

You can also generate from zero:

What do you hope to achieve in the final draft?

• Get straight to the point: Start with the subject, answer, or judgment, without letting the analysis plan and framework take the lead.
• Let specific people and events be subjects: Clearly state who did what, what happened, and what the result was, avoiding using abstract terms to replace real actions.
• Let judgments appear with evidence: Place facts, actions, numbers, or relationships close to judgments, without first declaring value before seeking support.
• Add new content to each paragraph: Only new facts, reasons, differences, consequences, or decisions count as progress; word substitution does not.
• Let the structure grow with the material: Organize content according to facts, cause and effect, timing, and questions readers will naturally ask, without pre-formulating fixed frameworks.
• Stop after finishing: Do not repeat previous texts for formality, forcibly summarize, elevate, or look forward.

What needs to be avoided

These are not forbidden words. Judge what a sentence is doing, then decide how to revise it.

• False Counterpart: Fabricating a misunderstanding out of thin air, then using “not … but …”, “seemingly … but actually …” to overturn the case and elevate the conclusion. When a real comparison exists, it should be retained and the specific differences clearly stated.
• Value Announcement: Use phrases like “truly important” and “worth noting” meaning “pre-declared value but without directly writing about changes and consequences.”
• Insight Signposts: Replace the real connection between facts, causes, and issues with prompts like “a deeper layer” or “a little more.”
• Process Narration: Leave the analysis and organizational processes such as “I organize first,” “expand below,” and “to ensure completeness” in the final draft.
• Abstract Actions: Use “optimization, enhancement, empowerment, closed loop” instead of specific subjects, actions, and outcomes.