Core Instruction Architecture

The core instructions are the second layer — loaded only when Claude has matched the front matter and determined the skill is relevant to the current task. This is where your SOP lives. The procedural steps, decision trees, quality checks, and output specifications that define how the skill actually executes.

The key design principle is numbered steps with explicit criteria. Not "research the topic thoroughly" but "Step 1: Identify the top 5 competitors by market share using web search. Step 2: For each competitor, extract pricing, key features, and target audience. Step 3: Create a comparison matrix with these columns: Company, Price, Key Differentiator, Target Segment, Weakness." Numbered steps give Claude a clear execution path. Explicit criteria give Claude a way to verify its own work at each step.

Core instructions should include decision logic — the if-then branches that handle variation. "If the CSV has headers, use them as column names. If the CSV has no headers, infer column names from the first 10 rows and ask the user to confirm before proceeding." Without decision logic, Claude encounters an edge case and improvises. Improvisation in a skill defeats the purpose of having the skill — you wrote it to ensure consistency, and consistency requires handling the branches.

Keep core instructions between 500 and 2,000 tokens. Below 500, the skill is too vague to be useful. Above 2,000, the skill is trying to do too many things and should be split into multiple skills. The token count is a design pressure that forces focus. If you cannot describe the skill's workflow in 2,000 tokens, the workflow is too complex for a single skill.