A lab for turning research into reusable Claude Code skills (and tools) for agentic development.
The loop: dump deep-research on an agentic-development topic, read it critically, then distill it into a self-contained, tested skill that improves day-to-day work with coding agents.
research/YYYYMM/<topic>-A.md # two independent research syntheses per topic
research/YYYYMM/<topic>-B.md
skills/<skill-name>/SKILL.md # skills distilled from the research
Every research topic has an A and a B synthesis from independent researchers. A tends to be cautious and evidence-grounded; B tends to be confident and prescriptive (and B's specific citations are sometimes unreliable — verify before trusting). Skills are designed from where the two agree.
| Skill | Distilled from | What it does |
|---|---|---|
solidify-repo |
agent-human-repo |
Audits a repo for human/agent collaboration — instruction files, deterministic checks, a one-command verify harness, and a three-layer security gate (SAST + dependency/supply-chain audit + secret scanning) — and applies fixes with per-category approval. Pulls stack-specific tool picks from a curated tooling reference. |
architecture-critic |
agent-driven-architecture-design |
Fresh-context adversarial review of a written design before code: a 7-dimension rubric, severity-tagged findings, and a PROCEED/REVISE verdict. |
feature-spec |
agent-driven-feature-spec |
Turns a vague feature idea into a right-sized spec before implementation — triage tiers to avoid over-production, fixed checklists for uniform coverage, and a machine-parseable handoff block for autonomous pipelines. |
autonomous-build-loop |
long-running-agents |
Conductor that drives a whole wishlist to verified completion in one long/unattended run — composes the other skills per phase, keeps state on disk for resumability, and scales pipeline depth to each feature's tier. |
Skills are authored and version-controlled here, but only take effect once copied into your global Claude Code skills directory:
cp -r skills/<skill-name> ~/.claude/skills/Then invoke it from Claude Code (e.g. "solidify this repo", or "review this design before I build it").
Each skill is authored with the superpowers plugin for Claude Code:
brainstorming to turn research into an approved design, then
writing-skills (test-driven — baseline the behavior with subagents, write the
skill to close the gap, refactor until solid).