You've been deploying this application for six months. The process has twenty-two steps. You could do it in your sleep — and that's precisely the problem. The knowledge lives in your muscle memory, not in any system that can share, teach, or automate it.
Skills extraction is the process of turning that implicit workflow knowledge into explicit, structured, reusable skills.
TL;DR
Skills extraction uses AI to observe your workflow sessions, identify repeatable patterns, and convert them into parameterized skills — structured representations of how you work that can be stored, searched, shared, and eventually automated. It bridges the gap between tacit knowledge (what you know how to do) and explicit knowledge (what's documented and transferable).
The tacit knowledge problem
Every organization has the same issue: critical processes live in people's heads.
The senior engineer knows the deploy sequence. The PM knows the reporting pipeline. The designer knows the handoff workflow. When any of them goes on vacation, the team slows down. When they leave, the knowledge walks out with them.
Traditional solutions — runbooks, wikis, SOPs — help but have a decay rate. They're accurate when written and increasingly wrong after that. Nobody updates the runbook because writing documentation is effortful and unrewarding.
Skills extraction flips the model. Instead of asking people to write down what they do, you observe what they do and extract the structure automatically.
How skills extraction works
Input: raw workflow data
The AI needs raw material to work with. Sources include:
- Screen recordings: visual record of the full workflow across all applications
- Terminal history: commands executed, in order
- Browser activity: URLs visited, forms submitted
- File system events: files created, modified, moved
- Application events: specific actions within tools (Figma frame exports, git commits)
The richer the input, the better the extraction. Screen recording captures the most complete picture because it includes cross-application workflows that no single event log can cover.
Processing: pattern recognition
The AI analyzes multiple sessions of similar work to identify:
Invariant steps — actions that happen every time in the same order. These form the skeleton of the skill.
Variable steps — actions where the specific value changes (a filename, a version number, a target environment) but the type of action stays the same. These become parameters.
Optional steps — actions that only happen sometimes, usually conditional on a previous step's outcome. These become branches in the skill.
Recovery steps — what you do when something goes wrong. These are some of the most valuable parts of a skill because they represent hard-won debugging knowledge.
Output: structured skills
A fully extracted skill contains:
name: weekly-metrics-report
description: Compile and distribute weekly product metrics
frequency: weekly (Mondays)
duration_estimate: 25 minutes
steps:
- name: Pull data from analytics
tool: amplitude
action: export_dashboard
parameters:
dashboard_id: product-weekly
date_range: "{last_monday} to {this_sunday}"
- name: Copy key metrics to spreadsheet
tool: google_sheets
action: update_range
parameters:
sheet: "Weekly Metrics"
tab: "{week_number}"
- name: Generate summary
tool: notion
action: create_page
template: metrics-summary
parameters:
metrics: "{extracted_from_previous_step}"
- name: Post to Slack
tool: slack
action: post_message
channel: "#product-metrics"
content: "{summary_link}"
variables:
last_monday: auto_calculated
this_sunday: auto_calculated
week_number: auto_calculated
error_handling:
- condition: "amplitude export times out"
action: "retry once, then use cached data from previous week"
- condition: "slack post fails"
action: "send via email to product-team@company.com"
This isn't a recording. It's a program derived from observed behavior. It can be reviewed, edited, shared with teammates, and eventually automated.
What makes extraction hard
The segmentation problem
When does one workflow end and another begin? If you check Slack in the middle of a deploy, is that part of the deploy workflow or a separate task? Context switches are frequent and rarely clean.
Good extraction algorithms use multiple signals: time gaps, application switches, semantic similarity of actions, and learned patterns from previous sessions.
The variable detection problem
In the URL https://staging.example.com/api/v2/users, which parts are fixed and which are variable? staging might change to production. v2 might change to v3. users is probably fixed. The AI needs multiple observations to distinguish variables from constants.
The intent inference problem
You clicked a button. Was that intentional, or did you misclick and immediately undo? You opened a file and closed it — were you checking something, or was it an accident? Extraction needs to separate signal from noise.
The cross-session alignment problem
Monday's deploy took 22 steps. Friday's took 19 because you skipped the staging verification (bad idea, but it happened). Aligning these sessions to extract the "true" workflow requires understanding which steps were skipped vs. which were genuinely different.
Practical applications
Onboarding acceleration. Extract the top 20 workflows from your senior engineers. New hires get a searchable skill library from day one instead of three months of tribal knowledge absorption.
Process auditing. Compare extracted skills across team members. You'll find that "the same workflow" is done differently by different people — sometimes with good reason, sometimes because nobody standardized it.
Automation planning. Before building automation, extract the skills first. You'll discover which workflows are stable enough to automate and which are still too variable.
Knowledge preservation. When someone leaves the team, their extracted skills stay. Not as stale documentation, but as structured, executable workflow descriptions.
Common mistakes
Extracting too early. Wait until a workflow is stable before extracting it. If the process changes every week, you're not extracting a skill — you're documenting chaos.
Treating extraction as one-and-done. Skills need maintenance. The best systems re-extract periodically and flag when the observed workflow has drifted from the stored skill.
Ignoring the human judgment parts. Not every step in a workflow is automatable. Some steps require human decision-making. The skill should clearly mark these as "human judgment required" rather than pretending they can be automated.
Skipping validation. An extracted skill is a hypothesis about your workflow. Have the person who does the workflow review the extraction and flag what's wrong. Unvalidated skills are unreliable skills.
How Distill does skills extraction
Distill captures your macOS screen sessions and runs extraction continuously in the background. Every time you complete a workflow, the system compares it against your skill library.
New pattern? It creates a draft skill. Existing pattern? It refines the stored skill with the new observation. Deviation from a known pattern? It flags the difference and asks whether the skill should be updated.
The result is a compounding skills library that grows more accurate and complete the more you work. No manual documentation required.
FAQ
How accurate is AI skills extraction?
After 3-5 observations of the same workflow, extraction accuracy is typically 85-90% for step identification and 70-80% for variable detection. Human review and correction improves this to 95%+. The first extraction is rough; the fifth is reliable.
Can I edit extracted skills manually?
Yes — and you should. Extraction produces a strong first draft, but human review catches edge cases, adds missing error handling, and corrects misidentified variables. Think of extraction as generating the skeleton and human editing as adding the muscle.
What's the difference between skills extraction and process mining?
Process mining works with system event logs (ERP transactions, database queries) and discovers processes at the organizational level. Skills extraction works with individual workflow observations (screen recordings, actions) and discovers processes at the personal level. They're complementary — process mining finds what happens, skills extraction finds how people make it happen.
How do extracted skills handle sensitive data?
Variables in extracted skills store the structure (this step requires a database password) not the value (the actual password). Sensitive data is identified and excluded during extraction. The skill knows a credential is needed at step 7; it doesn't store the credential.