Kaidera Skills Marketplace

Open-source skill manifests for AI agents, validated locally and published as a deterministic catalogue.

Kaidera Skills Marketplace is a repository of reusable skills for agents on the Kaidera Platform and for any harness that reads a SKILL.md. A skill is a Markdown instruction set with a security manifest. Every skill here is trust_tier: unvetted today: it has passed the repository's local checks and nothing more. What is implemented, and what is not says exactly what those checks are.

---

Start here: Kaidera Skills Catalogue and Operating Guide

Quick index: Kaidera Skills Glossary — one line per skill: what it does and why you reach for it

What Is a Skill?

A skill is one Markdown file, {skill-name}.SKILL.md, with a YAML frontmatter manifest followed by the instructions an agent reads:

• What it does (description, tags)
• Which tools it needs (kaidera.capabilities_required) and which domains it may reference (kaidera.allowed_domains)
• How risky it is (kaidera.risk_level) and how far it has been vetted (kaidera.trust_tier)
• Who wrote it, under which licence, building on whose work (author, license, attribution_*)
• What it must never do (safety_constraints)

Skills are instructions, not programs: the flat *.SKILL.md manifest carries no executable code. Two directory-form skills (gavel, jev) do bundle optional Python helper tooling under tools/, and that tooling is not covered by Gate 1 or Gate 2 — both gates read *.SKILL.md files only. See "Bundled tooling is outside both static gates" under Known portfolio debt in the catalogue. The full schema is in SKILL_FORMAT.md.

Skill Categories

Category	Description	Count
context/	Project and workspace awareness	2
development/	Code writing, review, testing	25
devops/	Deployment, infrastructure, CI/CD	3
documentation/	Specs, docs, changelogs, writing voice	1
legacy/	Retired EnGenAI-era material, kept as reference	21
research/	Research briefs, company research and evidence	2
sales-and-marketing/	Lead generation, outreach, social publishing and market listening	4
security/	Auditing, scanning, incident response	1

GSD and design selections

New portable source candidates: gsd-core, design-selection, ui-design-engineering and hyperframes. They are original adapters with pinned upstream citations; no upstream installer or plugin is bundled. The additional ui-ux-review and web-interface-guidelines candidates retain Apache-2.0/MIT donor terms and local notices. See review and selection boundaries.

The directory form can be installed through the universal plain-file skills path after project approval, for example: npx skills add Kaidera-AI/skills --skill design-selection -a universal -y --copy. Cortex projects register and bind through their existing owner; source publication does not activate these unvetted candidates.

What is implemented, and what is not

SKILL_FORMAT.md names four vetting gates. This repository implements the first two in bounded form and holds the other two. Nothing in this table is planned behaviour; it is what runs today.

Gate	Policy	State in this repository
1 Schema lint	YAML schema, required fields, forbidden patterns	Implemented (subset). node scripts/validate-skill.js --strict on every skill: strict YAML with duplicate-key rejection, required and unknown fields, name/filename and category/path agreement, real calendar dates, capability-to-risk minimums, the domain allowlist, 1 MiB and symlink limits. A formal JSON Schema is still future work.
2 Security scan	Injection and credential scan	Implemented (subset). The bounded regex set in spec/skill-security-patterns.json, applied by the validator and by python3 scripts/security-scan.py. It is not an LLM Guard scan and not a complete secret scanner.
3 Sandbox test	Execution in a capability sandbox	HOLD. No skill is executed anywhere. skill-sandbox-test.yml runs contract tests and a catalogue dry run only.
4 Human review	Reviewer approval, signing, provenance	HOLD. No reviewer root, no Cosign signing, no SLSA provenance. signed_by, last_reviewed and reviewer must stay empty; skill-publish.yml only reproduces the catalogue and fails on drift.

What follows from the holds:

• Validation accepts only trust_tier: unvetted. The trusted tiers (official, verified_partner, community_vetted) are rejected until a ratified Gate 4 verifier exists.
• unvetted skills are browsable and can be copied into any harness, but they cannot be bound to platform agents; a tier label or a catalogue row is not runtime evidence.
• Passing CI proves the named local checks. It is not sandbox acceptance, expert review, a signature, or publication approval.
• The Kaidera Platform's import queue, vetting reports and approval screens are not verified by this repository; see CONTRIBUTING.md, "Target Kaidera Platform Submission".

Also implemented: deterministic catalogue generation (npm run generate writes .claude-plugin/marketplace.json from the skill files; CI regenerates it and fails on drift), static routing evaluation for three skills (npm run eval:static, claim STATIC_CONTRACT_ONLY), and the catalogue tests that pin each skill's reviewed fingerprint (npm test). A skill projected from a canonical directory in another repository carries kaidera.source (repository, path, content hash) and is regenerated from that source, never edited here.

Using a skill

Copy the .SKILL.md file into the place your harness reads skills from. The body below the frontmatter is the instruction text; the frontmatter says what the skill needs and what it promises not to do. Read the catalogue entry first: it records each skill's operating posture (bounded candidate, reference-only, manual-only, rework before use).

kaidera-sdlc, gavel, and the five Jev-family skills also ship in directory form for the harness-agnostic skills CLI (replace the skill name as needed): npx skills add Kaidera-AI/skills --skill gavel -a universal -y --copy installs SKILL.md, the stdlib helper and its references into .agents/skills/gavel/.

For Jev, install jev and a wrapper into the same project separately: npx skills add Kaidera-AI/skills --skill jev -a universal -y --copy, then npx skills add Kaidera-AI/skills --skill jev-return-triage -a universal -y --copy. Each wrapper calls .agents/skills/jev/tools/jev.py; installing it alone does not install the core or the project's transfer policy. Use either Gavel or the three Jev lead wrappers (triage, handoff, backlog), not both. The Jev core uses TYPESAFE_API_KEY from the process environment. All remain unvetted; the CC-BY-4.0 root versus Apache-2.0 per-skill licensing precedence remains an explicit external-redistribution hold.

Skill Format

development-workflow is the generic AI+human profile of the SDLC framework: development-only routing, repository-off support, bounded tasks, independent code review, risk-based QA, recorded human decisions, and service operations. Its canonical directory is development-workflow; the flat manifest is generated from it. Select one lifecycle owner per project. Source availability does not lift the runtime or redistribution holds described above.

Every skill follows the SKILL_FORMAT.md specification. A minimal valid skill:

---
name: git-commit
version: 1.0.0
description: |
  Guides writing conventional commit messages following the Kaidera format.

kaidera:
  category: development
  trust_tier: unvetted
  risk_level: low
  capabilities_required: []
  allowed_domains: []
  content_hash: ""
  signed_by: ""
  last_reviewed: ""
  reviewer: ""

author: your-github-name
license: Apache-2.0
updated: 2026-09-04
tags: [git, commits, conventions]

safety_constraints:
  - Read-only reference. No tool access required.
---

# Git Commit

Write commit messages following: `type(scope): description`

Check it before opening a pull request:

node scripts/validate-skill.js --strict skills/development/git-commit.SKILL.md
python3 scripts/security-scan.py skills/development/git-commit.SKILL.md
npm test

Contributing

See CONTRIBUTING.md for the submission steps and the review process (strict lint, static scan, maintainer review, source merge).

Ways to contribute:

• Submit new skills via pull request
• Improve existing skills (bug fixes, better prompts)
• Report issues with skills that don't work as expected
• Suggest new skill categories

The repository root is licensed CC-BY-4.0 and each skill declares its own licence in its frontmatter; the precedence rule between the two is not ratified yet (see CONTRIBUTING.md, "Licensing"). You retain credit as the original author.

Attribution

This project uses the Creative Commons Attribution 4.0 International license. When we pull skills from external repositories, we:

1. Credit the original author in the skill's attribution_author field
2. Link back to the source repository in attribution_url
3. Preserve any additional license notes in attribution_notes

If you see your work here without proper credit, please open an issue and we'll fix it immediately.

Security

Two checks run on every skill, locally and in CI:

1. Schema lint (Gate 1 subset) — scripts/validate-skill.js: manifest shape, capability and domain declarations, forbidden patterns.
2. Static scan (Gate 2 subset) — scripts/security-scan.py and the shared pattern list in spec/skill-security-patterns.json: selected prompt-injection signatures, invisible Unicode, Base64 payload candidates, selected credential formats.

Not in place: sandboxed execution (Gate 3), human expert review with signing and provenance (Gate 4), and the runtime <skills_context> envelope, which belongs to the platform and is not verified here. Report a security problem in a skill by opening an issue; do not include exploit payloads in the report.

See SKILL_FORMAT.md, "Vetting Pipeline Target", for the full policy and its current status.

License

This repository is licensed under CC-BY-4.0 (Creative Commons Attribution 4.0 International).

You are free to:

• Share — copy and redistribute skills in any format
• Adapt — remix, transform, and build upon skills for any purpose

Under the following terms:

• Attribution — You must give appropriate credit to the original author

---

Kaidera - The Machine That Builds Machines

kaidera.ai | Platform Repo

The existing folder-organisation policy is mirrored at1.0.1 with original ownership-aware sync instructions and pinned MIT method attribution. It remains a manual-only, unvetted source candidate; source publication does not activate it.