Harness basics

What a Harness is

A Harness packages everything the project agent follows: roles, workflow steps and human review gates, rules, skills and templates for each deliverable. Each run takes a snapshot of the Harness when it starts, so later edits never change work already in progress.

• Product requirements (default for new projects): PRD and HTML prototype.
• Implementation delivery: technical mapping, implementation plan, and code drafts with tests.
• Game delivery: game brief, engineering plan, art direction and release checklist.
• Research & knowledge: source log, source comparison, and a research report with knowledge entries.

Three layers

• Built-in: the four Harnesses above. They are read-only.
• Organization: in the Harness library, create an organization Harness that extends a built-in one, then edit its files to add your own rules, skills and templates. Requires the admin role.
• Project: in the project’s Harness tab, switch rules and skills on or off, add skills and choose model tiers for that project only.

Layers merge in that order. An entry with the same id in a later layer replaces the earlier one, and built-in entries can be disabled but not deleted.

Skills

A skill is a short, reusable procedure with four parts: when to use it, the steps, how to verify the result, and known failure modes. Skills are injected into the agent’s context when their trigger steps or terms match.

The project Harness tab shows how often each skill led to an approved or rejected result, and its evaluation score.

Learning proposals

After a run finishes and after human review, the agent reflects on what worked and may propose project knowledge or a new skill. Each proposal includes its evidence.

1. 1Open Learning proposals in the sidebar.
2. 2Read the proposal and its evidence under Pending.
3. 3Approve it, or enter a reason and reject it.

Approved knowledge is added to the project. Approved skills are evaluated first and become active only if they pass. Rejected proposals are kept as negative samples, so the agent does not propose the same thing again.