Conducting Skill Extraction Interviews
When an input is too thin to generate a good skill, don't guess—ask. Present three targeted questions, one at a time, tracking progress:
Question 1 of 3: Your Process
Walk me through your step-by-step process. What do you do first,
second, third when working on this?
Collect all three answers, then merge them with the original input into an enriched context block (format below) and pass that to the skill generator.
Progress:
- Step 1: Assess if input scored below 60 (or is otherwise thin) — if so, trigger extraction
- Step 2: Ask Question 1 (Process) — validate answer is ≥30 characters
- Step 3: Ask Question 2 (Tools) — validate answer is ≥30 characters
- Step 4: Ask Question 3 (Real Example) — validate answer is ≥30 characters
- Step 5: Allow back/forward navigation between questions before finalizing
- Step 6: Merge original input + three answers into enriched context
- Step 7: Pass enriched context to skill generator
- Step 8: If user chooses to skip, use original input as-is and warn of likely lower grade
The Three Questions:
-
Your Process — "Walk me through your step-by-step process. What do you do first, second, third when working on this?" Captures the sequential workflow backbone.
-
Your Tools — "What specific tools, frameworks, or techniques do you use? Name the actual things you work with." Captures named, concrete tools/frameworks/techniques.
-
A Real Example — "Describe a real example: input, process, output. Walk through an actual case from start to finish." Gives a concrete case study for the skill's Examples section.
Enriched context format:
[Original input]
Step-by-Step Process
[Answer to Question 1]
Tools & Techniques
[Answer to Question 2]
Real Example
[Answer to Question 3]
Example 1: Input: "I'm good at recruiting engineers" (scores 35/100 — too vague) Process:
- Q1 answer: "First I analyze the requirements, then I research comparable solutions, next I create an initial draft..."
- Q2 answer: "I use the MATCH framework for screening, LinkedIn Recruiter for sourcing, and Greenhouse for tracking..."
- Q3 answer: "A startup needed a senior engineer. I sourced 50 candidates through LinkedIn and GitHub, screened 15 using MATCH criteria..."
Output: Enriched context combining all four pieces, sent to skill generator, producing a detailed
recruiting-technical-talentskill with concrete tooling and a worked example.
Example 2: Input: User clicks "Skip" instead of answering questions Output: Original thin input passed directly to generator as-is; resulting skill flagged as likely lower quality; user informed they can retry with more detail.
- Ask one question at a time; show a step indicator ("Question 1 of 3") so the user tracks progress.
- Enforce the 30-character minimum per answer — reject and re-prompt if shorter.
- Allow users to navigate back to revise earlier answers before submitting.
- Preserve the user's original input verbatim; append rather than replace.
- Encourage naming actual tools/frameworks (proper nouns) in Question 2 — specificity is what makes methodology actionable.
- For Question 3, insist on a complete arc (input → process → output), not just a description of a task.
- Don't skip straight to skill generation on thin input without offering the extraction flow first.
- Don't accept vague, generic answers (e.g., "I use good tools") — prompt for specifics.
- Don't discard the original input when answers come in — always merge, never replace.
- Don't force linear-only navigation — users need to revise earlier answers based on later reflection.
- Don't silently lower quality expectations when a user skips — clearly communicate the tradeoff.