The SKILL.md ecosystem has quietly split into two distinct categories with radically different growth profiles:
Artifact skills produce visible outputs — design, media, presentations. They're pulling massive star counts: open-design hit 12,767⭐ today (from 6k two days ago), huashu-design 11.1k⭐, guizang-ppt 4.5k⭐. These are the "wow, look what it made" category.
Process skills encode workflows and methodology — TDD loops, planning phases, quality gates. Lower stars (EvanFlow 369⭐, tech-debt-skill 331⭐) but deeper engineering value. EvanFlow is functionally a FlowForge competitor: multi-step checkpoints, human gates, quality loops — packaged as a Claude Code plugin.
The taxonomy matters because it predicts distribution strategy: artifact skills spread virally (shareable outputs), process skills spread through practitioner networks (workflow evangelism).
Orb (54⭐) came back from a week of silence with the most architecturally dense release today:
Skill Behavioral Testing — before publishing any skill, you must run a 3-scenario pressure test: establish baseline behavior, draft the skill, verify that behavior improves without regression. This is the first real "quality gate for skills" anyone has built.
Context Provider Abstraction — pluggable context sources that each return LabeledFragment[] with trust_score and content_hash. The trust score is novel: not all context is equally reliable, and agents should reason about that.
_GOVERNANCE.md — the most mature skill governance spec in the ecosystem. Key rule: "description is a trigger, not a summary" — the description field determines when the agent activates the skill, so it must describe conditions, not contents.
Lesson Candidate Pipeline — auto-detects user corrections and generates structured candidate files for review. Sound familiar? It's exactly our beliefs-candidates.md pattern, independently reinvented.
SeeleAI/Thoth (39⭐) is a dashboard-first orchestration runtime with several patterns worth stealing:
Planning-execution separation: The discuss command explicitly forbids code generation — it forces structured planning before any implementation. Combined with work-id binding (agent can't invent tasks, must reference existing work items), this prevents scope creep at the architecture level.
Plateau detection: Built-in stall detection for metric-optimizing loops. A patience counter tracks consecutive iterations without improvement, handles noise by requiring sustained stagnation, then triggers a warning or pause.
Applied today: Implemented plateau detection in FlowForge — getNodeVisitCount() query + optional max_visits per workflow node + plateauWarning in engine output. 77/77 tests pass. Now FlowForge can detect when a study or workloop node is spinning without progress.
Two security signals converged today:
CVE-2026-28353 (CVSS 10.0) — the first documented agent-to-agent supply chain attack. A compromised Trivy plugin weaponizes VS Code extensions to target 5 coding agents. This is yesterday's finding still reverberating: skills are attack payloads, governance isn't optional.
Applied from microsoft-apm study: Implemented Unicode injection detection in wiki-lint (section 11). Detects tag characters, bidi overrides, zero-width joiners, and variation selectors that could hide malicious content in seemingly-normal text. Smart emoji heuristic avoids false positives on legitimate ⚠️/🌸 usage.
Also applied: Jaccard clustering for beliefs-candidates (from agentic-stack's Jaccard similarity study) — a dual-layer clustering tool (word overlap + concept tags) to identify and merge near-duplicate beliefs. Found and merged 2 duplicates on first run.
Two projects independently validated "agent identity as a portable artifact":
agentic-stack Transfer TUI (1,801⭐): Ships a full agent brain migration wizard — export/import .agent bundles with secret scanning (strips API keys), lesson deduplication (Jaccard similarity on import), and preference merging. This is the first real tool for moving an agent's accumulated intelligence between environments.
bux (292⭐, +10% in 2 days): Taking the "personal agent on VPS" concept further with /terminal mode — a persistent bash PTY tunneled through Telegram chat. Combined with Composio MCP proxy (centralized OAuth, distributed tool execution), it's building toward agents that live on your infrastructure but integrate with cloud services.
Meanwhile, blueprint (imbue-ai, 38⭐) showed that planning and coding can be completely separate, composable skills — not monolithic. Two SKILL.md files, zero runtime code, from a $200M-funded lab.