Roadmap

source-to-skill is intentionally starting with a small core: local text, readiness scoring, and multi-level artifact generation.

The long-term direction is not "one source, one skill." The long-term direction is source-to-delta: new sources should become evidence, refinements, contradictions, seeds, or new skills depending on how they relate to the existing skill system.

v0.1: Readiness Gate

Status: shipped.

• Local UTF-8 text and Markdown input.
• Human-readable readiness reports.
• JSON readiness reports for integrations.
• Note, Skill Seed, Mini Skill, and Full Skill artifact builders.
• Basic SVG identity and documentation.

v0.2: Intake Plugins

Goal: normalize more source types into text before scoring.

• Local HTML intake. Shipped.
• Remote HTML / article intake. Shipped.
• Transcript cleanup command. Shipped.
• EPUB intake. Shipped.
• PDF intake.
• Keep extraction separate from scoring.

v0.3: Skill Quality Evaluation

Goal: make generated skills easier to trust.

• Smoke-question runner.
• Evidence coverage checks. Shipped.
• Claims-without-evidence warnings.
• Fold-in quality report for existing skills.

v0.4: Audio And Long-Form Sources

Goal: support audio without making the product "recording-to-skill."

• Local transcript intake first.
• Optional Whisper CLI transcription wrapper. Shipped.
• Transcript cleanup and topic splitting. Shipped.
• Default audio output should usually be Note or Skill Seed unless the score is strong.

v1.0: Stable Skill Compiler

Goal: a small reliable tool that can be used in real agent workflows.

• Stable CLI contract.
• Stable JSON schema.
• Better documented scoring.
• Bundled end-to-end demo. Shipped.
• Format adapters for common agent skill layouts.
• Real examples from books, articles, interviews, and meetings.

v1.1: Skill Evolution Layer

Goal: prevent skill sprawl by updating existing skills before creating new ones. See docs/review-gates.md for the first-principles and adversarial review model this layer should use.

• Skill metadata scanner for existing skill folders.
• Source-to-skill matching by domain, title, use case, and evidence overlap.
• Relationship classifier:
  • duplicate
  • evidence
  • refinement
  • contradiction
  • new skill
• Pending update artifacts under evolution/pending-updates/.
• Human review flow for contradictions and major rewrites.
• First-principles review:
  • core problem
  • core method
  • use case
  • boundary
• Adversarial review:
  • counterexamples
  • contradictions
  • overgeneralization
  • missing evidence
  • unsafe merge risk
• Changelog entries for accepted updates.
• Re-run evidence and regression checks after a skill evolves.

v1.2: Skill Router Inputs

Goal: make large skill libraries easier for agents to use.

• Generate or update skill metadata:
  • use_when
  • do_not_use_when
  • domains
  • trigger signals
  • maturity
  • confidence
• Create a lightweight skill index for routing.
• Recommend the smallest relevant skill set for a user task.
• Avoid loading every skill into the agent context.