A recipe app lets cooks scale ingredient quantities. Halving servings produces the wrong flour amount. Its developer needs a repeatable test-and-review repair process; preserving a product specification alone does not establish that habit.
How: work toward a solution
Agree on the scaling rule, write a test for 200 g becoming 100 g, observe failure, fix the calculation, and rerun checks. Review the change against the agreed plan.
A recipe scaling test goes from observed failure to a checked repair
Recipe app · illustrated test-first repair
Observe failure. Repair. Review.
Agreed behavior
200 g × 2 / 4 = 100 gFour servings → two servings
01 / Failing test
✕ 200 g
Observed result differs from the expected 100 g
02
Fix the scaling calculation
→
03 / Passing test
✓ 100 g
Rerun the test and relevant checks
04 / Review against the plan and check other agreed quantities
Illustrated procedure, not an executed test. Harness setup and skill activation must be checked; existing project limits still apply.
What: the concept
Superpowers packages development workflows as composable agent skills. Installation and activation depend on the harness. Its procedures add work; prefer existing checks when sufficient, and a specification workflow when durable requirement tracking is the priority.
The priority here is a repeatable development procedure: clarify design, plan work, verify a regression, and review the result. Superpowers includes skills for those activities. Spec Kit or OpenSpec can hold durable requirements if needed; identify which records and approvals govern execution before combining them. Existing project instructions and an ordinary agent may already provide enough discipline. These selection criteria are editorial judgments, not measured quality rankings.
02Applications
The fictional recipe uses 200 g of flour for four servings. Two servings require 100 g; an illustrated failing result of 200 g exposes the bug. A meaningful test exercises the scaling behavior, then the developer observes it fail before changing code and pass afterward. Review includes other agreed quantities and rounding rules where relevant. The diagram illustrates this process and does not execute tests or demonstrate an actual model result.
03Implementation & cautions
Read the current official installation instructions for the external agent in use. The project documents plugin or extension installation by harness; Codex CLI currently exposes a plugin selection flow. If the assistant cannot operate that flow, provide the exact user step and retain the setup status as pending. Inspect the installed skills, hooks, scope, and active version before use. Invoke or load the needed skill and confirm its output; installation alone does not establish activation. Reuse approved requirements and preserve project-specific test and delegation limits. Do not treat subagent, worktree, merge, or publish instructions as extra authorization. A full workflow can impose more ceremony than a small correction needs.
Already talking with external AI? Share this article link and explain why the choice fits your project and which tradeoff matters. Ask how it applies to your agreed requirements.
Discuss this idea
Comments are shared across languages. GitHub login required to post.