Roles & Skills
How agents adopt roles, get graded on skills, and improve over time.
Why roles matter
Pulse AI uses roles to keep agent work from collapsing into generic assistance. A role tells the agent what kind of judgment to apply, which standards matter, and what evidence a reviewer should expect.
The same task can fail or pass depending on role fit. A frontend task needs UI judgment. A documentation task needs factual writing and information architecture. A dispatch task needs operational discipline.
Role adoption flow
Before work starts, the agent should read the task, choose the closest role, and apply that role's guidance. The role should match the outcome, not the agent's personal default.
For example:
- Content and education work belongs with a content or education role.
- UI layout and frontend behavior belongs with a frontend role.
- Dispatch state, board hygiene, and workflow trust belongs with an operations role.
- Package, source, and runtime wiring belongs with an engineering role.
If the task pivots, the role should pivot too. That keeps the review standard aligned with the real work.
Skills and rubrics
Roles point agents toward the skills and rubrics they need. Skills describe how to do the work. Rubrics describe how the work will be judged.
Rubric corrections are especially important. They preserve user feedback from earlier runs so agents do not repeat the same mistake. If a correction says not to widen scope, not to publish internal checklists, or not to invent docs content, that correction is part of the current quality bar.
Reviewing role fit
A reviewer should ask:
- Did the assigned role match the outcome?
- Did the agent use the right skills for the work?
- Did the result satisfy the role-specific standard, not just a generic checklist?
- Did the agent capture new corrections when the founder gave explicit feedback?
If the answer is no, the task should move back for correction or receive a specialist review before approval.
Common routing mistakes
- Defaulting every task to engineering when the real risk is content, product, or operations.
- Treating review as a style pass instead of an evidence check.
- Letting a broad task keep expanding after the role boundary is clear.
- Ignoring rubric corrections because the current task feels urgent.
Source of truth
Role files live under .pulse/org/. Skill guidance lives under .agents/skills/. Rubric corrections live under .pulse/behavior/skills/rubrics/.
When these sources disagree with public docs, source-backed workspace guidance wins and this page should be updated.