From Prompting to Mature AI Workflows: The 20% LLM Principle
Build AI skills that package context and process, then move deterministic checks (formatting, linting, tests) outside the LLM to create a robust 20% LLM workflow.
Refactor your AI workflow to separate skills from gates; implement automated checks for formatting, linting, and tests, and keep the LLM only for ambiguous tasks.
Summary
Prompting is the discovery phase of an AI workflow, where the model helps shape ambiguous tasks by generating drafts, examples, and constraints. As prompts grow longer, they become fragile because the model must balance style, facts, tone, and platform conventions all at once.
The next stage is to package this context into reusable skills that reduce repetition, but skills can balloon into prompt monsters if they absorb too many rules. A mature workflow separates the LLM’s role—handling ambiguity—from deterministic gates such as formatting, linting, tests, and schema validation, resulting in a 20% LLM workflow that keeps the model focused on the parts it can best handle. For content, gates include title‑promise matching, section flow, banned phrases, and link verification; for code, gates include formatting, linting, type checks, unit tests, and CI checks. The article emphasizes that the model should not be the final authority on whether output passes, but rather a tool that drafts and suggests while the system enforces standards. By moving deterministic checks outside the LLM, teams can build robust, repeatable AI workflows that scale with minimal reliance on the model’s mood.
Key changes
- Prompting is the discovery phase of an AI workflow, where the model helps shape ambiguous tasks by generating drafts, examples, and constraints
- Longer prompts become fragile because the model must balance style, facts, tone, and platform conventions all at once
- Reusable skills package context and process to reduce repetition, but can balloon into prompt monsters if they absorb too many rules
- A mature workflow separates the LLM’s role—handling ambiguity—from deterministic gates such as formatting, linting, tests, and schema validation, resulting in a 20% LLM workflow
- For content, gates include title‑promise matching, section flow, banned phrases, and link verification; for code, gates include formatting, linting, type checks, unit tests, and CI checks
- The article emphasizes that the model should not be the final authority on whether output passes, but rather a tool that drafts and suggests while the system enforces standards
- By moving deterministic checks outside the LLM, teams can build robust, repeatable AI workflows that scale with minimal reliance on the model’s mood
- Teams should refactor AI workflows to separate skills from gates, implement automated checks for formatting, linting, and tests, and keep the LLM only for ambiguous tasks