-
-
Logo
Inspiration
I kept doing deep research with researchkit and then losing the know-how — a report is not reusable behavior. I wanted every research pack to become something an agent can do: an installable skill, with the claims verified and the triggers tested. That is skillskit. During the hackathon window alone I used it to author and publish a whole shelf of skills — the live catalog is at https://www.skills.sh/paldom.
What it does
skillskit is the one-shot path from research to reusable agent behavior. Its core skill, /skill-from-research, takes a research pack (a researchkit run, or any folder of reports and notes), inventories it, verifies load-bearing claims against primary sources on the web, splits the material into single-purpose skills, and authors each one eval-first: every skill ships with should-trigger and should-not-trigger evals plus quality checks before it is allowed to exist. /create-skill-repo scaffolds a new skills repository, /publish-repo ships it to skills.sh, and /add-skill covers authoring from an idea when no research exists.
The output installs anywhere: npx skills add Paldom/<repo> targets 70+ agents — Codex, Claude Code, Cursor, Copilot and the rest — and working inside a checkout, Codex discovers the same skills directly via the Agent Skills standard (.agents/skills/).
How we built it
skillskit is agent-native by design — the pipeline is skills, prompts, validators, and evals rather than a compiled program, because eval-first authoring needs a model in the loop. A deterministic validator gates every skill (frontmatter contract, trigger-overlap checks, eval completeness), and CI runs the same discovery as skills.sh so a broken skill won't ship.
Codex and GPT-5.6 were working reviewers throughout: the researchkit-pack contract that /skill-from-research consumes was cross-reviewed by a multi-vendor panel that included GPT-5.6 Sol at its highest reasoning tier, and the skills produced during the hackathon were exercised in Codex sessions as a first-class target (the .agents/skills/ discovery path was verified live with codex exec). Research packs themselves are produced with GPT-5.6 as a provider and meta-summarizer inside researchkit — so the knowledge a skill encodes has passed through GPT-5.6 both as a research provider and as a design reviewer before it ships.
Challenges we ran into
Research packs go stale, and agents love to skim one file and invent the rest. The pipeline fights both: a pack inventory step that flags empty and duplicate files before hours are wasted, mandatory verification of versions and APIs against primary sources, and a hard rule that committed skills cite source URLs — never gitignored research paths. Scoping was the other fight: if a skill needs "and" to describe, it gets split.
What we learned
Eval-first is the difference between a prompt file and a skill. Writing the should-NOT-trigger cases first exposed how many skills would have collided; the validator now checks trigger overlap across a whole repo's catalog before anything merges.
What's next
Tighter pack-contract validation against researchkit's versioned spec, richer eval tooling, and more of the catalog: everything I research from now on either compounds into a brainkit brain or ships as a skill.
Log in or sign up for Devpost to join the conversation.