Where Do Agent Capabilities Come From? A High‑Frequency Skill Toolset for AI Agents

Agent Skills are custom command scripts that extend an AI agent's native text‑only capabilities by invoking local tools such as Python, Shell, SQL, diagram generators, document converters, HTTP clients, and image processors.

Python

Python is used to quickly build HTTP services, batch‑process Excel files, and send network requests. Running a script is as simple as:

python 脚本名.py

Scenario: Operations need to merge 12 department Excel files with inconsistent formats. Prompt to AI: “I have a folder with 12 Excel files, each with a header row. Write a Python script that merges them into one file, removes duplicate headers, and saves as merged.xlsx.” AI generates the script, which reduces a half‑day task to a 10‑second execution.

Shell (Bash)

Shell scripts automate system‑level tasks such as log archiving. Example prompt:

Scenario: Developers must archive log files older than 30 days. Prompt to AI: “Write a Shell script that finds .log files older than 30 days in /var/log/app/, packs them into log_archive_YYYYMMDD.tar.gz, deletes the originals, and adds the script to a daily 2 am cron job.” AI returns the script and cron configuration, enabling one‑click automation.

SQL

Complex SQL functions can be generated on demand. Example prompt:

Scenario: Operations need 7‑day retention rates per channel from MySQL tables users and events. Prompt to AI: “Write a SQL query that counts registrations per channel in the last 90 days and the number of users who logged in within 7 days, then compute retention rates ordered by highest.” AI supplies a complete JOIN, date calculation, aggregation, and ordering statement ready for execution in Navicat.

Mermaid

Mermaid describes diagrams in plain text and renders instantly in VS Code. Example prompt:

Scenario: A product manager needs a sequence diagram for a purchase flow involving user, front‑end, back‑end, and payment system. Prompt to AI: “Generate Mermaid code for the sequence: user clicks purchase → front‑end submits order → back‑end creates order → front‑end triggers payment → payment system processes and callbacks → back‑end updates status → front‑end shows success.” AI returns the Mermaid source, which renders the diagram with a single click.

Pandoc

Pandoc converts between dozens of document formats. Command pattern:

pandoc input_file -o output_file

Scenario: A Java team has a README.md that must be turned into a styled Word document for a client. Prompt to AI: “Write a Pandoc command that converts README.md to Word, generates a table of contents, sets A4 page size, uses SimSun font, preserves heading levels, and applies a reference.docx style template.” AI provides the full command sequence, producing a ready‑to‑send document.

curl / API Calls

curl commands replace Postman for quick API testing and scheduled data pulls. Example workflow:

Scenario: Operations need to fetch yesterday’s revenue report via an internal API and push it to a WeChat Work webhook each morning at 8 am. Step 1: Prompt AI to create a curl command for GET https://api.internal.com/report/daily?date={yesterday} with an Authorization header. Step 2: Prompt AI to generate a second curl command that sends the JSON field total_revenue to the webhook as a markdown message. AI returns both commands, which can be combined into a Shell script and scheduled with cron.

ImageMagick

ImageMagick batch‑processes images, adding watermarks, resizing, and converting formats. Installation commands:

brew install imagemagick

sudo apt install imagemagick

Scenario: E‑commerce operations receive 300 product images that need to be cropped to 800×800, watermarked, and converted to WebP. Prompt AI: “Write an ImageMagick command to resize all JPGs in the current directory to 800×800 with white padding, output to resized/.” Prompt AI: “Write a command to add logo.png as a bottom‑right watermark 20 px from the edge, output to watermarked/.” Prompt AI: “Convert all JPGs in watermarked/ to WebP at quality 85, output to final/.” AI supplies three commands that process all images in one run.

PlantUML

PlantUML uses plain text to generate UML diagrams, stored as .puml files in Git. Example prompt:

Scenario: A technical proposal requires a component diagram of a micro‑service architecture (API Gateway, User Service, Order Service, Product Service, Message Queue, MySQL, Redis). Prompt to AI: “Generate PlantUML code for a component diagram with the listed services and their interactions.” AI returns the .puml source, which can be rendered by a plugin, version‑controlled, and updated by editing the text.

These tools together cover roughly 80 % of everyday automation scenarios. By encapsulating recurring prompts as reusable Agent Skills, teams can let AI agents execute concrete commands, turning knowledge of “what to ask” into “how to act”.