Skills Over Prompts
Superpowers and the Engineering Discipline Paradigm for AI Coding

Jesse Vincent, creator of Superpowers (author of Request Tracker and Perl 5 release manager), sharply summarized the failure modes of AI coding agents: "Language models are trained to be 'immediately helpful,' an instinct that drives them to produce code as quickly as possible while skipping all prerequisite engineering steps."

The result is: wrong libraries used, incorrect file locations, missing key requirements, and ultimately humans spending more time debugging the agent's output. Vincent's solution does not modify the model itself; instead, it intercepts this default instinct at the session level and routes it through a structured gatekeeping process—a design choice that treats "engineering discipline" as a first-class citizen.

The Superpowers repository is organized into three layers, each addressing different problems:

The most disruptive design is "Markdown as Specification": using Markdown instead of JSON/YAML to encode process specifications makes the same file serve as both human documentation and machine instructions. Frontmatter provides metadata, the body describes behavioral constraints, embedded Graphviz DOT diagrams visualize workflows, and "Red Flags Tables" list common AI rationalizations with rebuttals—a Single Source of Truth.

When a user says "Let's build X" in a conversation, Superpowers does not immediately write code but proceeds through seven steps—each step is a mandatory gate, not a suggestion:

1. ① brainstorming · Socratic design refinement, asking only one clarifying question at a time and presenting solutions in segments.
2. ② using-git-worktrees · Creating isolated workspaces and verifying clean test baselines.
3. ③ writing-plans · Breaking designs into 2-5 minute micro-tasks with precise file paths and complete code.
4. ④ subagent-driven-development · Dispatching a fresh sub-agent for each task to avoid long-session drift.
5. ⑤ test-driven-development · "NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST"—ironclad wording.
6. ⑥ requesting-code-review · Performing pre-reviews between tasks and categorizing issues by severity.
7. ⑦ finishing-a-development-branch · Verifying tests, merging/PR/discarding, and cleaning up workspaces.

Note step ④—a new sub-agent is dispatched for every task. Behind this lies a profound observation: agents in long sessions gradually deviate from original specifications, and the longer the context window, the easier it is to "forget" initial design decisions. Sub-agent isolation is not for parallel acceleration but to combat attention drift through context resets.

Anti-Rationalization: AI tends to "make excuses" when facing constraints, much like humans. Superpowers specifically lists common evasion excuses in the using-superpowers skill ("I've already tested manually," "This is too simple to need design") along with pre-written rebuttals. writing-skills even references Cialdini's persuasion principles from "Influence," teaching skill authors to leverage psychological mechanisms like authority and consistency to strengthen constraints.

Testing Skills, Not Code: Superpowers' testing system does not verify code correctness but validates whether skills effectively prevent agent violations. Test scripts design over 3 combined stress scenarios to force agent violations, verbatim recording specific excuses—this "meta-testing" approach is unique among agentic frameworks.

Process Gates Over Prompt Engineering: Rather than investing in more sophisticated prompts to "persuade" agents to do the right thing, it is better to implement mandatory gates at the architectural level. Superpowers proves that when agents have no choice but to follow specifications, behavioral quality improves significantly.

• Three-Layer Separation Model: Decoupling behavioral specifications, injection mechanisms, and platform adaptation allows one set of specs to serve multiple tools.
• Markdown as Executable Specification: Making specifications simultaneously human documentation and machine instructions reduces maintenance costs.
• Process Gates Over Prompt Engineering: The optimal point for constraining agent behavior is the architectural layer, not the prompt layer.
• Sub-agents as Context Management Tools: Using context resets to combat attention drift in long sessions.
• Testing Skills Instead of Code: Redefining "correctness" in the context of AI agents.

If you are designing AI coding or research workflows for your team, we invite you to explore ResearchLinkAI's ARS Academic Research Skills Library—we have implemented the industry's latest engineering discipline methodologies across three output lines: research papers, patents, and software copyrights.