Skills

Skills are instruction sets injected into the system prompt. They tune the model's behavior without adding new tools or capabilities. superclaw has 5 built-in skills.

How skills work

A skill is a block of instruction text. When a skill is activated, its instructions are appended to the base system prompt before the first model call. The model sees the instructions as part of its context and adjusts its behavior accordingly.

Skills do not add new tools. They cannot make the agent access resources it couldn't access before. They are a way to bias the model's reasoning and output format for a particular class of task — nothing more.

base system prompt

skill instructions (appended)

system_prompt field (appended last)

Skills overview

Skill	Best for	Output bias
summarize	Distilling long documents	Concise bullets, headers
research	Web-based information gathering	Multi-source, cited
extract	Structured data from unstructured text	JSON / Markdown table
coding	Software development tasks	Explained code, test-first
github	GitHub issues, PRs, changelogs	GH-flavored Markdown

summarize

skill

Instructs the agent to produce concise, structured summaries. Emphasizes brevity, key points, and avoidance of filler. Best for distilling long documents, changelogs, or research into readable output.

When to use

When the task is primarily about reading and condensing content — README summaries, changelog digests, meeting notes, article TLDRs.

Prompt behavior

Adds instructions to: produce a summary no longer than 20% of the source, use headers and bullets, skip boilerplate and legal disclaimers, and cite the source file.

research

skill

Optimizes the agent for web-based research tasks. Biases it toward fetch_url and web_search calls, cross-referencing multiple sources, and expressing uncertainty where sources conflict.

When to use

When the task requires gathering information from the web — comparing libraries, checking latest versions, finding documentation, or surveying a topic.

Prompt behavior

Adds instructions to: always fetch at least 2 sources before concluding, include source URLs in the output, note where sources disagree, and avoid stating unverified claims as facts.

extract

skill

Instructs the agent to extract structured data from unstructured content. Favors JSON or table output. Good for pulling structured information from HTML, text files, or API responses.

When to use

When the task is to extract structured data — scraping a table, pulling fields from a document, parsing a log file for events, or converting freeform text to JSON.

Prompt behavior

Adds instructions to: return extracted data as JSON or Markdown table, include a confidence score when extraction is ambiguous, and never invent values that weren't present in the source.

coding

skill

Tunes the agent for software development tasks. Emphasizes reading existing code before editing, writing tests, using idiomatic patterns, and explaining changes.

When to use

When the task involves writing, reviewing, or debugging code — implementing a feature, fixing a bug, refactoring, generating tests, or explaining a codebase.

Prompt behavior

Adds instructions to: always read existing files before modifying them, prefer patch_file over write_file for edits, run tests after changes with run_bash, include a brief explanation of each change.

github

skill

Specializes the agent for GitHub-related workflows: reading issues, summarizing PRs, generating changelogs, and writing release notes. Works well with fetch_url pointed at GitHub URLs.

When to use

When the task involves GitHub content — summarizing an issue thread, writing a PR description, generating release notes from commits, or reviewing a diff.

Prompt behavior

Adds instructions to: parse GitHub issue and PR URLs directly with fetch_url, format output in GitHub-flavored Markdown, include issue/PR numbers in references, and use conventional commit conventions for changelogs.

Activating skills

Via CLI flag

bash

# Single skill
superclaw -skills summarize "summarize CHANGELOG.md into 5 bullet points"

# Multiple skills (comma-separated)
superclaw -skills coding,github "review the open PRs and write a summary"

Via superclaw.json

superclaw.jsonjson

{
  "skills": ["research", "extract"]
}

Combining skills with a custom system prompt

Use the system_prompt field to add project-specific context after the skill instructions:

superclaw.jsonjson

{
  "skills": ["coding"],
  "system_prompt": "This is a Go project using the standard library. Prefer table-driven tests."
}

Combining skills

Multiple skills can be active at once. Their instruction blocks are concatenated in the order listed. There is no conflict resolution — if two skills give contradictory instructions, the model uses its judgment.

Effective combinations:

research + summarizeGather information from the web and produce a concise summary.
coding + githubReview a PR diff and generate a structured review comment.
extract + summarizePull structured data and then distill it into a readable report.