oberskills

oberskills

Skills that make Claude Code better at the things it's worst at: writing like a person, searching the web without hallucinating URLs, building skills that actually work, and not embarrassing itself when dispatching agents.

As of v2.0.0 the three meta-skills — prompt, agent, skill-craft — are full skills (skills/<name>/SKILL.md) rebuilt on a 2026 research pass (Anthropic platform docs, ~100 arXiv papers, practitioner practice), and the skill-eval pipeline is a Bun/TypeScript MCP server instead of Python scripts.

Skills

prompt

Claude-first prompt engineering. DESIGN mode drafts or fixes a prompt; REVIEW mode audits one adversarially and returns a verdict table. Nine evidence-cited principles inline; eight reference files on demand (Claude model deltas, verbatim Anthropic snippet library, context engineering, optimization, security, porting to non-Claude models). Covers the current-era shifts: de-prompting, prefill removal, adaptive thinking and effort instead of CoT incantations, purpose-conditioned few-shot counts, reasoning-before-answer schema ordering.

agent

Subagent dispatch guidance, built for the moment you write an Agent call: dispatch-vs-inline cost gate, the four-part delegation contract, model and effort selection for the current lineup, fan-out sizing, fork vs fresh subagent, and debiased verifier dispatch (no intent framing, separate verifier, weaker model allowed). References carry platform mechanics, orchestration patterns, and the canonical verification-bias evidence.

skill-craft

Create, evaluate, and review Claude Code skills. CREATE runs intake → design → baseline → build → eval → ship with gates; REVIEW audits a skill directory. Judgment stays with Claude; everything checkable runs through the skill-eval MCP tools.

skill-eval (MCP server)

Bun/TS server bundled with the plugin (mcp/), spawning real headless Claude sessions via the Agent SDK. Seven tools:

Tool	Does
validate_skill	Frontmatter/structure/content lints (agentskills.io spec + house rules), optional .skill packaging
test_triggers	Live trigger-rate measurement: does the description actually route?
optimize_description	Iterative description optimization with a held-out test split (chunked: one iteration per call)
run_eval	Run one eval with/without the skill, optional pressure blocks (composed in code, 3+ enforced), auto-grade
grade_run	Externally-dispatched grader; severities and verdicts computed in TypeScript, not by the LLM
aggregate_benchmark	Mean/stddev/min/max + deltas across configurations, gate evaluation
compare_outputs	Blind A/B comparison of two outputs

Dependencies install automatically via a SessionStart hook into ${CLAUDE_PLUGIN_DATA} (requires bun on PATH). After install or update, run /reload-plugins once. Optional: allow mcp__plugin_oberskills_skill-eval__* in settings to skip permission prompts.

write

Two modes. EDIT rewrites silently. REVIEW walks you through issues one batch at a time, asks questions, then offers an edit pass. Built on 47 AI-writing detection papers, Pangram Labs data (N=millions), and a blind test that dropped AI detection probability from 85% to 15%.

web-research

Parallel search agents fan out across multiple dimensions (docs, tutorials, discussions, forums). Each agent extracts precise information with source URLs. Results synthesize back through your model. No hallucinated links.

shot

Screenshot capture and analysis. Full screen, active window, or named window. Dispatches a haiku-tier analyzer and returns a summary.

clarify

Decomposes user intent through structured brainstorming before acting on ambiguous requests. Model-invoked; other skills chain into it.

How They Connect

skill-craft ──┬─ CREATE: intake → design → baseline → build → eval → ship
              │           └── skill-eval MCP tools (validate, triggers, evals, grading)
              └─ REVIEW: validator floor → quality dimensions → behavioral test

prompt ──┬─ DESIGN: principles + on-demand references
         └─ REVIEW: adversarial audit → verdict table
              (owns ALL prompt review, including agent prompt files)

agent ──── dispatch gate → delegation contract → model/effort → verifier dispatch
              (chains to prompt for long/novel briefs, clarify for ambiguous intent)

write ──┬─ EDIT: core rules + surface rules (+ deep craft if needed)
        └─ REVIEW: scan → orient → top issues → next batch → offer edit

web-research ─── parallel search agents ─── synthesize with source URLs
shot ──── capture → haiku analyzer → summary

Install

/plugin marketplace add ryanthedev/rtd-claude-inn
/plugin install oberskills@rtd
/plugin update oberskills@rtd

Then /reload-plugins (or restart) so the skill-eval MCP server connects.

Version

2.0.0

---

MIT