Skills are the atomic operations that tools are built from. Each step in a tool invokes one skill. Skills range from deterministic text operations to AI processing, external API calls, public webpage retrieval, and trusted local Python scripts.
What a Skill Is
Each skill has:
Name — A unique identifier such as ProcessPromptWithAI, FilterLines, or CallConnectedAPI.
Description — What the skill does, written so the planner knows when to use it.
Parameters — The inputs it accepts.
Tags — Category labels such as ai, file, memory, text, or data.
Skill Categories
Interaction — User prompts, confirmations, chat responses. Most are system-only. ChatResponse is the main exception.
AI — Model-powered processing. ProcessPromptWithAI sends instructions to an AI model. ProcessWithThinkForgeMCP lets the model use only the local ThinkForge MCP operations explicitly allowed for that step. CombineDocuments merges multiple inputs using AI.
Text — Deterministic text manipulation: extract regex matches, find and replace, filter lines, sort, deduplicate. Fast, no AI call, predictable output.
Data — Structured data operations: parse CSV to JSON, query JSON by path, filter rows, select columns, aggregate, render as Markdown tables.
File System — File operations: copy, move, delete, get info, search. Most are system-only.
Project — Project-scoped file operations. ReadFile and WriteFile are the most commonly used.
Validation — Input checking. ValidateInput is available to tools; email, path, URL, JSON, and safety checks are system-only.
Memory — SK Memory search: search conversations, find similar discussions, get summaries, analyze patterns.
Mind Map — Create, expand, classify, and generate code from mind map nodes.
Web and Browser — FetchWebPage retrieves text, metadata, or server HTML from public webpages without opening Chrome or running JavaScript. Interactive browser automation is not implemented.
API — External service calls through CallConnectedAPI, which reaches 11 services.
Local Process — RunPythonScript runs one trusted, user-selected Python script through an explicitly selected interpreter.
Headless Web Retrieval
FetchWebPage can process one or more public HTTP or HTTPS addresses. It returns page text by default, with optional HTML or metadata-only modes. A bare address uses the protocol selected in Tool Settings; HTTPS is the default. You can set a maximum of 1 to 25 URLs and a per-page text limit.
This is static-page retrieval, not browser control. It does not use signed-in browser sessions, execute JavaScript, access private networks, or bypass logins, paywalls, CAPTCHAs, or bot protection. Requests are time-limited, redirects are revalidated, and downloads are size-limited.
Available vs System-Only
Not all skills are available to user-created tools. The Capability Manifest controls what the creation planner can use.
Available to tools — ProcessPromptWithAI, ProcessWithThinkForgeMCP, CombineDocuments, ChatResponse, all Text skills, all Data skills, ReadFile, WriteFile, ListFiles, GetFileInfo, SearchFiles, ValidateInput, all Memory skills, all Mind Map skills, CallConnectedAPI, FetchWebPage, and RunPythonScript.
System-only — Interaction skills such as AskUser, ConfirmAction, and ShowMessage; file management such as Copy, Move, and Delete; all prompt skills; and interactive browser skills, which are not implemented.
The distinction prevents user-created tools from performing dangerous operations such as file deletion, or operations that only make sense during tool creation.
Skill Tiers
Each skill has a tier indicating its impact:
- Tier 0, read-only — Search, retrieve, validate. No side effects.
- Tier 1, safe write — AI processing, file writes. Produces output but stays inside the project.
- Tier 2, controlled side effects — API calls, webpage retrieval, local process execution, and MCP-backed AI processing. An MCP-backed step can affect project data only through its explicit tool allow-list.
- Tier 3, admin — Not available to user tools.
How Skills Receive Data
Each step receives its saved arguments, settings, and a structured envelope containing the current data. Skills that require a URL, path, or another specific value extract that value from the envelope themselves. The runtime does not guess or globally bind values between steps.
- Static arguments from the tool definition
- The previous step's result
- The original input
- Tool instructions, which AI skills use
- Tool settings, which configurable skills use
- Source and context information
Composing Skills into Tools
A tool is a sequence of skill calls. The planner picks the right skills and puts them in the right order:
- Data retrieval skills go early — webpage fetches, API calls, file reads, memory searches
- Processing skills go in the middle — AI, text manipulation, data transformation
- Output skills go last —
WriteFile,ChatResponse,ToMarkdownTable
The runtime chains them automatically. You do not need to plan data flow between steps.