tacit skills extraction
How We Extracted 10 Hidden Skills from 5945 AI Prompts
A real experiment: 784MB of Claude Code sessions analyzed with heuristics to surface tacit human skills invisible to traditional assessment.
What are tacit skills and why they matter
Tacit skills are the invisible expertise you use every day but cannot list on a resume. Scope management, diplomatic communication, architectural instinct — these emerge from how you work, not what you claim. Traditional assessments miss them entirely.
The experiment: 784MB of real AI sessions
We analyzed 5945 prompts from 78 Claude Code sessions spanning a full product build. No synthetic data, no curated examples — raw interaction logs from months of building a production system for a *******-backed startup.
Five heuristic extractors, zero LLM cost
Instead of running each prompt through another AI, we built pattern-based extractors: correction detection (when the human overrides the AI), tool frequency analysis, decision language parsing, domain clustering by file paths, and prompt structure classification. Each extractor produces typed signals with confidence scores.
Results: 5 skills the developer did not know he had
The analysis surfaced Scope Killing (expert-level judgment for cutting features), Spec-Driven Architecture (methodology for spec-first development), Duplication Audit (systematic code dedup), Diplomatic Communication (reframing feedback constructively), and Visual Prototyping (designing UI through iterative specs). None were in any resume or self-assessment.
From discovery to verifiable proof
Each extracted skill was ingested into a Temet agent card with structured evidence, hash-chained receipts, and public verification. The skills are not self-declared labels — they are machine-observed patterns backed by cryptographic proof of when and how they were recorded.
Try it yourself
The session analyzer is open source. Point it at your own Claude Code sessions and discover what skills you have been demonstrating without realizing it. Then turn those signals into a public skill profile with structured evidence.
Thanks for reading. Arnaud
FAQ
Does this require sending my prompts to an external AI?
No. The heuristic extractors run locally on your machine. No prompts leave your system.
How accurate are heuristic-extracted skills?
Heuristics catch patterns, not intent. They surface candidates that a human should review. In our experiment, 5 out of 10 candidates were genuinely surprising and validated by the developer.
Can I use this with other AI assistants besides Claude?
The session parser currently reads Claude Code JSONL format. Other formats require adapting the parser, but the heuristic extractors are format-agnostic.
Next step
Use this guide in practice with Temet's audit, tracking, and profile workflow.
Connect your agentPublished February 12, 2026