Agent Skills
Assessing
The kinds of skills SpecStory mines, and the confidence signals that help you decide which to keep.
SpecStory mines two kinds of skills from your sessions. Knowing which is which helps you read the confidence signals and decide what to keep.
Kinds of skills
Theme skills (latent)
A theme skill captures a latent practice — a way of working that shows up across your sessions without ever being explicitly named. SpecStory detects these by clustering similar conversational patterns and measuring whether they correlate with better outcomes.
Theme skills are badged latent in the library. They tend to be the more valuable (and more surprising) finds: "you consistently resolve merge conflicts faster when you summarize the diff out loud first" is a theme skill.
Correlation skills (corr)
A correlation skill captures a repeated command or action pattern — a workflow you actually executed that the miner noticed you do more than once. These are closer to traditional "runbook" skills: a sequence of steps that worked.
Confidence
Every skill carries signals that help you assess how much confidence to place in it. You'll see these in the expanded row on both the web page and the CLI.
| Signal | What it means |
|---|---|
| Confidence | The miner's confidence in the candidate (e.g. high / medium / low). |
| Judged cross-vendor | The skill was validated across multiple agents (Claude, Codex, Cursor, Gemini, Droid), not just one — so it generalizes. |
| Sessions of evidence | How many of your synced sessions contributed to this candidate. More is better. |
| Prevalence (theme) | How often the pattern appears across your sessions. |
| Outcome lift (theme) | Whether sessions where the pattern appears had better outcomes — measured in points above baseline. +12 pts outcome lift means sessions with this pattern scored 12 points higher than your baseline. |
How to read outcome lift
Outcome lift is the strongest confidence signal for theme skills. A positive lift (+N pts) means the practice correlates with better outcomes; a negative lift means it correlates with worse ones (and the miner is less likely to have confirmed it). Lift is only computed when there are at least 3 judged members, so a missing lift field just means there wasn't enough data yet.
Note
Correlation is not causation. These signals describe what tends to happen in your sessions — they're decision aids, not guarantees. Your judgment on Keep/Dismiss is the final filter.
The review decision
When a skill is in Review, you're the human filter. A quick read:
- High confidence + cross-vendor + multiple sessions of evidence → likely a solid Keep.
- Low confidence or single-session evidence → read the
SKILL.mdand decide; it may be early or may be noise. - Negative outcome lift → probably Dismiss, unless you recognize the pattern as something you're deliberately working against.
You can always change your mind later — approved skills can be uninstalled, and the miner will re-surface patterns as your session history grows.
Provenance
Every skill records the cluster key it was mined from — a stable identity for the cluster of sessions that produced it. This is mostly diagnostic, but it lets you trace a skill back to the conversations that generated it if you want to understand why the miner proposed it.