hexiaofeier/sha-hao-yong

Sha-hao-yong (啥好用)

Writes a pre-installation manual for candidate skills that even a non-technical person can understand, answering: What is this? Is it good? Is it right for me? Final deliverable is a single-file HTML.

Default is read-only. Does not install, run unfamiliar code, read credentials, register, pay, upload, or modify the candidate.

Reference Routing

• Judgment ability comes from: Read complexity-map.md.
• Calculate any quantitative score: Read scoring-rubric.md first, then the corresponding indicator file.
• Calculate installation complexity: Read installation-score.md.
• Calculate task complexity: Read task-complexity.md.
• Assess skill invocation accuracy: Read invocation-accuracy.md.
• Calculate token consumption: Read token-score.md.
• Calculate risk score: Read risk-score.md.
• Calculate recommendation fit: Read fit-assessment.md.
• Write user-visible text: Read writing-style.md.
• Output HTML: Read report-template.md, and strictly use scripts/render_report.py.

Only read the References needed for the current task.

Workflow

1. Lock Assessment Target

• Clearly identify whether you are assessing a single Skill, an entire repository, a plugin, a CLI, an application, or a product entry point.
• If a repository contains multiple objects, describe them separately. Do not apply repository-level metrics directly to a single Skill.
• When you need to display a name to the user, take the clearest item from the information already provided or immediately visible: page or page title, last segment of the URL, folder name, file name, or the name field in the target file's YAML. If unobtainable, display 【】. Do not read the body, dependencies, or installation instructions prematurely to identify the name.

2. Choose Assessment Method

After launch, if the user has not explicitly chosen an assessment method, display only one multiple-choice question:

• A: Evaluate based on my needs
• B: Evaluate based on the general target audience

If interactive controls are supported, use a clickable single-choice component. If not, display the text options and ask the user to reply with A or B. Stop after displaying, and wait for the user's choice. Do not read candidate materials, collect evidence, or begin assessment prematurely.

• If the user chooses B: Proceed directly to step 3. Define the general target audience based on the candidate object's official positioning. Clearly state the judged object in the report. Do not ask further about requirements.
• If the user chooses A: Using the name identified in step 1, ask one mandatory open-ended question: You want to evaluate the skill [Target Skill Name]. What do you hope it can help you do? Stop after displaying, and wait for the user's answer. Once you receive the answer, proceed immediately to step 3. Do not ask further about device, AI agent, cost, tokens, privacy, or operational preferences.

If the user has already explicitly chosen A or B in the startup message, go directly to the corresponding branch. Do not repeat the first question. If the user chooses A and has already stated their specific needs, proceed directly to step 3. Do not repeat the question. Do not treat environment or preferences the user did not provide as facts.

3. Collect Evidence

Prioritize checking the target file, official documentation, installation instructions, dependency list, release notes, license, and official pricing. Only check leaderboards, Stars, Forks, update time, Issues, tests, and real-world cases when you need to assess popularity and maintenance status.

• Use the current AI agent's existing file reading, web access, or search capabilities. Do not require the user to install additional tools.
• Prioritize checking materials provided by the user and official sources. Write unverified for information that cannot be found. Do not guess.
• If you cannot access the internet or read links, continue the assessment with the materials you have. Only ask the user to provide core files if you cannot determine what the target object is.
• Only display "Downloads" in the report if verifiable public download numbers are found. If no numbers are available, delete that line directly.
• Do not use Stars, Forks, or collection repository data to represent the usage effect of a single Skill.
• A repository being publicly readable or downloadable does not mean it is free to use or permits commercial use. Do not write "core is free" if the license is not confirmed.
• Write unverified for pricing, compatibility, and maintenance status if not found. Do not guess.

4. Answer Three Questions

What is this Skill? Only answers:

1. Who is it suitable for, under what circumstances should they invoke it, and what problem does it solve?
2. What does the user need to provide, and through what method and core steps does the Skill complete the task?
3. What is the final deliverable? Can the result be used, modified, or verified directly?
4. Who or what actually does the work: the AI agent, computer software, an online service, or a combination?

Is this Skill good? Only answers:

1. Public attention: Verifiable Stars, Forks, leaderboards, and public download numbers (if any).
2. Installation complexity: Software, dependencies, accounts, keys, permissions, system compatibility, and ongoing maintenance.
3. Task complexity: Input preparation, execution steps, branching, manual confirmation, rework, and failure recovery.
4. Skill invocation accuracy: Based only on name, description, and trigger rules, assess whether it is easy to miss-invoke or mis-invoke.
5. Verifiability: Are documentation, maintenance, updates, Issues, tests, and real-world cases sufficient?
6. Additional costs involved: Software subscriptions, APIs, online services, and other payment points.
7. Token consumption: The model budget required to complete a typical core task.
8. Risk score: Data, permissions, external operations, and potential loss involved.

Is this Skill right for you? Only answers:

1. Whether it matches the task the user wants to complete and the result they want to achieve.
2. Which scenarios are suitable and which are not.
3. Whether installation is recommended.

Do not repeat the content from the first two parts. Only put "Usage rhythm suggestions" for first trial, expansion, review, and deactivation methods.

5. Unified Scale

• Installation complexity: 0—10. Higher scores mean more effort for installation and maintenance.
• Task complexity: 0—10. Higher scores mean more input, operations, judgments, and manual handling needed to complete one typical task.
• Skill invocation accuracy: Qualitative judgment as Accurate, Too Broad, Too Narrow, or Cannot Assess. Clearly state how the name, description, and trigger rules affect invocation.
• Token consumption: 0—10. Higher scores generally mean more model budget consumed.
• Risk score: 0—10. Higher scores mean more data, permissions, and external operations involved.
• Recommendation fit: 0—10. Higher scores mean it is more worthy for the current user or the audience identified in the report to adopt.

One decimal place is allowed. Each score must be accompanied by its level, main reason, and basis for judgment. The user report must not contain S5, P1, A20, F0, or a five-item recommendation breakdown.

6. Output Fixed HTML

First, locate the directory where the current SKILL.md resides, then call scripts/render_report.py under that directory. Do not assume the user's current working directory is the Skill directory.

First run init to generate the complete JSON skeleton, then fill it in according to report-template.md. After completion, perform a short-sentence quality check against writing-style.md.

<Python command> <Skill directory>/scripts/render_report.py init <Skill name> --output-dir <delivery directory>

After filling in the JSON, verify and generate in one command:

<Python command> <Skill directory>/scripts/render_report.py build <process JSON path>

<Python command> uses the actual available Python 3 on the device: Common on Windows is py -3 or python; common on macOS/Linux is python3. Claude Code can use ${CLAUDE_SKILL_DIR} to locate the Skill; other AI agents directly use the directory where the SKILL.md was read.

The build command lists all validation issues at once. Upon passing, it automatically generates the HTML to the parent directory of process file_CO/, and checks for fixed sections, a 10-block recommendation bar, footer notes, preparation instructions, gradients, and placeholder fields. To only check, run <Python command> <Skill directory>/scripts/render_report.py check <process JSON path>.

The script only uses the Python standard library. It fixes JSON fields, the Obsidian Lacquer layout, the 10-block recommendation bar, the conclusion card, and the disclaimer.

• Do not write HTML or CSS by hand.
• Do not manually modify the report body after rendering.
• If the script reports an error, modify the JSON. Do not modify the script to bypass the check.
• The JSON is placed by init into process file_CO/ under the delivery directory. The final deliverable only points to the HTML.

Writing Constraints

• Break the Skill positioning into 3—4 short sentences. Each sentence should explicitly use the target Skill name, the user, or "you" as the subject. Do not start with "This is" or "This is a set of."
• The first sentence of the Skill positioning at the top of the page should state the target Skill name. When referring to it again in the second and third sentences, use "it." The fourth sentence may restate the name for closure.
• For other cards or paragraphs, the first sentence should state the target Skill name. When referring to it again later, use "it." Keep "user" and "you" as intended. If another object that could cause confusion appears, restate the target Skill name.
• Use parallel phrases for "Suitable scenarios / Unsuitable scenarios." Do not apply the subject rule. Do not repeat "Suitable" or "Unsuitable."
• The order of the conclusion card is fixed: Final Conclusion, Quantitative Analysis, Applicable Scenarios, Usage Recommendations.
• The first sentence of the final conclusion must include the target Skill name, e.g., "auto-tutor partially matches your needs."
• When judging by applicable user, rephrase to "Suitable for this type of user," "Partially suitable for this type of user," or "Not suitable for this type of user." Who specifically this refers to must be clearly written next to the recommendation bar.
• "Is this Skill right for you?" should be pinned at the top showing the matching conclusion and installation recommendation. The renderer ensures both are consistent with the overall assessment.
• The quantitative analysis section only reports key numbers. The applicable scenarios section only writes who it is suitable for, what they will do, and what they will get. The usage recommendations section directly writes whether to install or not and when to consider it.
• When usage recommendations are linked with the final conclusion, the final conclusion first writes the target Skill name, and the usage recommendations use "it" to follow. If recommending conditionally, the specific condition must be stated in the first sentence.
• "Usage timing, user provides, task steps, additional needs, final deliverable" should all be short bullet lists, each item not exceeding 35 Chinese characters (note for English: keep concise).
• The recommendation bar should be divided into 10 equal-width color blocks, displaying the score from light to dark. The scoring description should be placed below the color bar.
• Prefer tables, lists, and short sentences. Write each fact only once.
• Do not use internal terms like "host," "execution chain," "static scan," "evidence type" in the user report. Rephrase them as "AI agent," "actual steps," "only checked public files," "basis for judgment."
• Prohibit unexplained abbreviations like multi_agent, CI, LLM, PR. Consistently use "AI agent." Do not use "multi-agent," "sub-agent," or "AI programming agent."
• State known conditions positively. Prohibit reverse educational sentence patterns like "Don't understand... as..." or "It's not... but..."
• Prohibit indirect or roundabout expressions. If you can write "It cannot complete meeting minutes," do not write "Core work still needs to be completed by other capabilities."
• Allow a small amount of relaxed expression, but do not use jokes for cost, risk, privacy, and the final conclusion.
• Do not use popularity as a substitute for effectiveness, simple installation as a substitute for quality, or low risk as a substitute for security certification.
• If the basis for scoring is insufficient, provide a range or write unverified. Do not fill with zeros or fake precision.

Boundaries

• This is a pre-installation manual, providing only adoption reference. It cannot replace professional evaluation, including code security audits, malicious code detection, software quality and compatibility testing, privacy compliance review, license review, or legal advice.
• Read-only inspection must not be passed off as actual installation testing, compatibility testing, performance testing, or long-term usage experience.
• Do not ask the user to paste API Keys, Cookies, Tokens, passwords, or other credentials into the conversation.
• The report must remind the user to independently verify the source, dependencies, permissions, data destination, fees, and license before downloading, installing, authorizing, paying, uploading, publishing, or running automated operations. Prioritize using a test environment, least privilege, dedicated accounts, and recoverable backups.
• Scores are only for adoption reference. The project version, pricing, platform rules, and dependencies should be re-evaluated after updates.
• If sources conflict, retain them side-by-side and mark them with [Needs Verification].
• The footer must explain "AI agent": An AI assistant capable of reading Skills, understanding tasks, and invoking computer software or online services to complete work according to rules. Specific permissions depend on the product and user authorization.