Test whether an AI skill earns its keep.
Would reusable instructions make your AI better at a task you've already completed? Turn that work into a repeatable test, then compare several kinds of skill with the same coding agent working without one. See which approach improves the result, how much time and effort it takes, and whether any improvement is clear enough to matter.
From Feedforward's Praxis. Created by Ethan Mollick, Feedforward.
Use it on Praxis
This move runs as a ready-made skill in Praxis, Feedforward's skill library.
How it goes
What the AI will do
- Recover the original request, starting files and finished output, then turn them into a repeatable task with concrete checks and a quality rubric.
- Build a few relevant skill variants, such as short instructions, a detailed workflow or instructions with helper scripts, and compare them with the agent using no skill.
- Run each approach several times with the same agent and model, grade the results, and report quality, consistency, time and token use where available.
What to have ready
- A task you completed with a coding agent, such as Claude Code or Codex, and access to its session or files.
- The inputs before the task and the finished work to use as a reference, if they can't be recovered from the session.
- What a successful result should accomplish and how much testing you want to do.
Where you come in
- Review the task, pass/fail checks and quality criteria before approving the comparison.
- Confirm the number of variants and runs, with their estimated cost, before the trials begin.
- Examine the results and decide whether to adopt a skill, refine one or keep using the agent without it.
What you'll get
A reusable benchmark, the outputs and scores from each run, and a leaderboard comparing the skill variants with the plain agent. You also get a recommendation grounded in the measured differences and, when a variant wins, that skill ready to install. A finding that no skill clearly helps is a valid result.
You're done when
A reusable benchmark, the outputs and scores from each run, and a leaderboard comparing the skill variants with the plain agent. You also get a recommendation grounded in the measured differences and, when a variant wins, that skill ready to install. A finding that no skill clearly helps is a valid result.
*Praxis is available to Feedforward members.