Encoding Senior Engineering Discipline into AI Agents

AI coding agents often act like juniors by skipping critical engineering steps like testing and documentation in favor of speed. 'Agent Skills' is a framework of markdown-based workflows that forces these agents to follow senior-level SDLC practices and provide verifiable evidence of their work. By encoding these rigorous standards into the agent's context, developers can prevent the technical debt and incidents typically caused by unvetted AI code.

Key Points

• AI agents default to junior behavior by focusing on code output while skipping essential non-coding tasks like specs and reviews.
• Effective agent skills must be structured as actionable workflows with specific checkpoints and exit criteria rather than reference essays.
• Anti-rationalization tables are used to pre-emptively counter the plausible excuses LLMs generate to bypass engineering rigor.
• The framework utilizes progressive disclosure to manage context window limits, loading only the skills relevant to the current task phase.
• Senior engineering discipline, such as scope control and non-negotiable verification, must be explicitly encoded into agent prompts to ensure software reliability.

Sentiment

The community is notably divided. A substantial camp of experienced practitioners defends the approach, sharing concrete production success stories and arguing it represents a natural evolution of software engineering practices. An equally vocal skeptical camp questions both the reliability of LLM instruction-following and the rigor of productivity claims. The middle ground — that these tools work as force multipliers for experts but are far from autonomous — appears to have the broadest implicit support, with even strong proponents acknowledging that domain expertise and human oversight remain essential.

In Agreement

• Structured workflows and skills genuinely improve AI output quality, similar to how organizational processes improve human output — the key is building reliable systems from unreliable components
• Agent Skills and similar frameworks encode valuable engineering discipline that prevents AI from taking shortcuts, and experienced developers report significant measurable productivity gains in production environments
• Skills are most effective when treated as reusable context snippets that guide AI toward project-specific conventions, API usage patterns, and non-obvious requirements rather than as rigid rule enforcement
• The process-over-prose philosophy is sound — actionable workflows with clear exit criteria outperform passive documentation for both humans and AI agents
• The anti-rationalization table concept addresses a real problem where LLMs generate plausible excuses for skipping quality steps, and having pre-emptive rebuttals helps maintain discipline

Opposed

• LLMs are fundamentally non-deterministic and will inevitably drop hard requirements from markdown files, making skill-based approaches unreliable — human review remains the only trustworthy quality gate
• These skill frameworks are cargo-culted Rube Goldberg machines that haven't been rigorously A/B tested, and similar outcomes could be achieved with a simple prompt asking the agent to plan first and ask questions
• The productivity gains from AI coding are overstated or illusory — developers are building features nobody asked for, and the true cost including debugging and refactoring time is rarely measured
• Encoding engineering expertise into AI-readable formats risks making developers obsolete by handing over their means of production to employers who may decide they no longer need human engineers
• The frameworks burn excessive tokens and context window space without proven benefit over vanilla prompting, and they will likely be deprecated as LLMs internalize these patterns through training