Harness Skill Generator for Openclaw

Prerequisites

• An agent environment that loads skills from a shared skills directory.
• A clearly defined workflow with enough complexity to benefit from phases, branching, quality checks, or persistent state.
• User availability for architecture, reference, sample-test, and delivery confirmations.

Scaffold a new skill

Use the generated skill name in place of <skill-name>:

mkdir -p <skill-name>/{references,templates}

Create the initial files:

touch <skill-name>/SKILL.md
mkdir -p <skill-name>/references <skill-name>/templates

Populate SKILL.md with frontmatter containing the skill name, description, and trigger phrases. Add the generated phase workflow, boundaries, checkpoints, quality protocol, and core rules. Then add one reference file per phase, plus optional files such as quality-checklist.md, style-contract.md, branch-routing.md, and repair-rules.md.

Install the completed skill

After testing and user approval, copy the complete directory to the shared agent skill location:

cp -R <skill-name> /Users/ii/.agents/skills/<skill-name>/

Update AGENTS.md with the skill entry when required, and record the creation event in MEMORY.md. This generator creates skills; it does not execute the newly created skill or modify and upgrade an existing skill.

Generated skill package

Planning and state artifacts

• plan/problem-definition.md stores the problem statement and pain points.
• plan/human-workflow.md records the manually performed workflow with inputs, outputs, and failure risks.
• plan/solution-research.md compares mature market solutions, agent-native approaches, and simulated human interaction options.
• plan/agent-workflow.md defines phases, dependencies, branches, outputs, quality checks, checkpoints, and state files.
• Project-level state files persist important decisions and enable recovery after interruption.
• Output files, state files, scripts, and review documents remain separated rather than scattered in the project root.

Reference metadata taxonomy

Each reference should define: what the phase does, how to perform it, input and output formats, anti-patterns, and a three-to-five-item self-checklist. Quality controls may be inline self-review, SubAgent message-based review, SubAgent file-based review, or user acceptance, depending on risk.