From one skill to the whole workflow.
Test whether AI gets the job done, stays within its limits and knows when to stop. The task sets the rules, not the model brand.
- Skills and instructions
- Does the agent follow the instructions, produce the required work and handle unusual requests?
- Agent loops
- Does it make progress, recover from errors and stop instead of repeating or spending without a limit?
- Workflows and teams of agents
- Do tool calls and handoffs produce the right result while respecting permissions and human review?
- The system around the agent
- Do required checks run, records stay intact and release rules stop work that is not ready?
Published examples currently focus on skills and the platform's own checks. Testing your workflow means connecting its runner and recording the steps you want checked. See current tooling and integration limits.
Real results, explained
Read the short version, or inspect these published studies. A result applies to the setup and checks described, not every possible use of the system.
-
A real comparison: quick rewrite vs. structured improvement draft
We gave the same rewriting task to a quick, single prompt and to our step-by-step Skill Refiner process. See both outputs, why one was chosen, what it cost, and the record behind the result.
-
We test our own published skills draft
Two AI troubleshooting guides, two different release decisions. One passed; the other was blocked after it missed a required report. The study records its checks, skipped judging and limitations alongside the results.
Detailed test definitions
These pages describe the checks behind the studies. Active means the test is in use; draft means it is under review. A definition's status is not a test outcome.
-
Agent skills: the seven-check framework active
Checks for structure, activation, task completion, repeatability and release readiness. The current Jrig command runs five checks by default; two more are opt-in. Read each run's coverage before interpreting its result.
Found an unclear rule? Each definition links to its source on GitHub so you can raise an issue.