01
Capture only on request
A person asks StyleSeed to preserve an accepted correction. There is no background project scan and no automatic observation mode.
$ss-learn captureOptional repository extension · not in the core install
Learn from human judgment. Not private projects.
StyleSeed can preserve a correction that a person already accepted, turn it into a bounded local candidate, and prepare one reviewed local package. Capture, sharing, optional bridge exposure, and core-rule promotion remain separate decisions.
02
Generalize the lesson
The candidate records applicability, counterexamples, measured evidence, and limits—not the project code, screenshot, prompt, or brand material that produced it.
local candidate03
Review separately
A caller-attested decision accepts or rejects the immutable draft. Preparing a share package requires another explicit attestation.
$ss-learn review04
Grant one exact read
The prepared package stays local and untransmitted. The extension's development-only bridge remains disabled until a host-owned proof adapter is verified; enabling it would expose one exact package to the connected client and model after a one-time grant.
one package · one grantAllowed candidate material
Generalized design judgment
- Applicability and explicit non-applicability
- Measured evidence and confidence limits
- Counterexamples and regression expectations
- Content hash and immutable review history
Rejected raw material
Project surveillance
- Source code, private prompts, screenshots, or URLs
- Names, reviewer identity, and local file paths
- Brand tokens, credentials, or arbitrary extra fields
- Background crawling or automatic core-rule edits
Promotion boundary
A candidate is evidence. Never authority.
One successful correction may be local and accidental. StyleSeed keeps it useful without pretending it is universal. Promotion requires repetition across contexts, named review, counterexamples, and regression proof.
Does StyleSeed train on my project?
No. The current workflow creates a local, generalized candidate only after an explicit request. It does not crawl repositories, upload raw material, or train a hosted model.
Does an accepted candidate become a StyleSeed rule?
No. A share package is evidence, not a rule. Team or core promotion still needs repeated cross-project evidence, counterexamples, regression coverage, and maintainer approval.
Is the MCP result private from the model?
No. Today the prepared package stays local and untransmitted, and the repository-only bridge must remain disabled. If a verified host adapter later enables it, the exact approved package would become visible to the connected client and model after the one-time grant is consumed.