YYLO Skills
Reusable agent skills maintained by YYLO. This repository is the canonical, independently versioned source for skills used by YYLO CLI, YYLO Ledger and YYLO Benchmark.
skills.sh Mentioned in Claude Skill Registry Mentioned in Agent Skill Exchange Mentioned in GAIA Skill Tree Mentioned in Claude Skills Collection Mentioned in Awesome Skills Mentioned in SkillsAllYouNeed Featured in PM Skills Community
Install
Install interactively with the open skills CLI:
npx skills add yylo-dev/yylo-skills
Install all skills for Claude Code, Codex, and Pi without prompts:
npx skills add yylo-dev/yylo-skills --skill '*' -a claude-code -a codex -a pi --copy -y
Install one skill by user-intent-first name:
npx skills add yylo-dev/yylo-skills --skill wiki-yylo
Every canonical slug ends in -yylo, keeping selector prefixes useful: type w for wiki/workflow, a for artifacts, p for planning, r for the Ralph loop, or u for project understanding. Review skill instructions and scripts before installing them. Published versions use immutable vMAJOR.MINOR.PATCH tags; VERSION identifies the repository release represented by the default branch.
Claude Code plugin installation
The repository also exposes one plugin containing all eight canonical skills:
claude plugin marketplace add yylo-dev/yylo-skills
claude plugin install yylo-skills@yylo-skills
Restart Claude Code after installation. Skills are namespaced by the plugin; use Claude's skill discovery rather than assuming the unqualified names. To validate an unpublished checkout without changing your normal settings, use an isolated configuration directory and a local marketplace path:
export CLAUDE_CONFIG_DIR="$(mktemp -d)"
claude plugin validate .
claude plugin marketplace add "$PWD"
claude plugin install yylo-skills@yylo-skills
claude plugin list --json
Installation supplies instructions, not the YYLO executables or a configured
controller. The four Ledger Record skills require a compatible yylo-ledger
(or yy ledger); project planning/execution require YYLO CLI and its controller;
benchmark-yylo requires YYLO Benchmark. Inspect installed command help before
use. Installation never authorizes task execution, publication or deployment.
Do not install another copy over customized skills; choose one acquisition route.
Maintainers keep the marketplace entry version and description aligned with
.claude-plugin/plugin.json and VERSION; node scripts/validate.mjs rejects
drift. Local source validation is not a public release or marketplace acceptance.
Additional plugin formats
Cursor metadata lives in .cursor-plugin/plugin.json; portable Hermes/Agent
Plugins metadata lives in root plugin.json. The Nix flake builds the portable
layout through agent-pkgs. DSH uses package.json's bundle declaration and a
small provider under dsh/. Grok Build has generated .grok-plugin/plugin.json
metadata; Grok Bot runtime compatibility remains unverified. All formats
reference the same eight skills. See host plugin packaging for installation,
runtime prerequisites, regeneration and the host acceptance checklist. A manifest
is not proof of marketplace acceptance or end-to-end host compatibility.
Native delivery and compatible CLI installation
Version 2.1.1 extends the existing benchmark-yylo skill with frozen project/task
checklists, evidence-based judgments, deterministic loss and reusable examples.
It keeps skills in this dedicated repository rather than the Benchmark npm package.
CLI 0.2.11 declares skills ^2.1.1 and Benchmark 0.2.1. An immutable v2.1.1 tag
is the skills release boundary; a version edit alone is not publication or an
installed upgrade. Existing immutable tags and release evidence remain unchanged.
Version 2.1.0 source updates benchmark-yylo for Benchmark 0.2.0's breaking
thin-runner lifecycle: reviewed case, run, independent evaluate, report and
append-only disqualify. It documents workflow-prefix comparisons, interchangeable
harnesses and trusted-host hygiene without claiming a sandbox or mandatory pilot
ceremony. CLI 0.2.10 source requires skills ^2.1.0. Source declarations alone do not
establish publication or an installed upgrade; inspect immutable release tags.
Version 2.0.4 introduced benchmark-yylo, completing the eight-skill set required
by CLI 0.2.9, whose source declared skills ^2.0.4. Version 2.1.0 replaces that
skill's old Benchmark lifecycle guidance; inspect installed Benchmark help/version
before use. Existing 0.1.x study evidence retains its original interpretation.
Version 2.0.3 requires user-facing Record results to include the actual kind/profile, immutable ID, and Ledger-returned slug. Missing fields must be reported as unavailable, not guessed. Validator regressions protect this contract across all canonical skills.
Version 2.0.2 clarified optional read-only preflight, finish-enforced admission
and validation, explicit project tests/reviews outside merge, and native
yy merge status|land|project delivery with projection recovery. The canonical
Ralph implementation contract lives in skills/ralph-loop-yylo/references/implement.md;
edit it here, not in installed agent directories or old CLI template copies.
The current source contract places implementation, tests and commits with the external agent. The CLI prepares workspaces and verifies delivery. Autonomous task execution and budget recovery are retired; optional watch observation never launches, retries, cancels or completes work. Explicit continuation must preserve the workspace, historical evidence and current ownership checks. These source changes require a separately reviewed immutable skills release; source commits do not update already-installed skills or authorize publication.
A compatible YYLO CLI declares its supported skill range in yyloSkills.version.
Run yy skills install or yy skills update explicitly to select the latest
compatible stable release and update all three agent copies. Unchanged
receipt-owned skills upgrade without force; differing customized/unrecorded
copies require review. Unrecorded legacy copies remain preserved and flagged,
not silently deleted. CLI installation alone does not replace skills.
Validate before delivery with node scripts/validate.mjs. It checks nested
references for retired lifecycle instructions as well as invocation contracts
and plugin version parity. Run node --test scripts/validate.test.mjs for regression
coverage. Maintainers publish each new version with a new immutable tag; never
replace an existing release tag. capabilities.json
retains evidence for previously published releases; do not invent new release
hashes before publication.
Universal Ledger retrieval guidance
The source skills teach yy ledger get ID (yylo-ledger get ID standalone) as
the universal read, with task_, doc_, and artifact_ storage-kind prefixes for
new generated IDs and unchanged historical identities. PDR purpose does not
require a PDR prefix: new PDRs remain artifact/report, with historical
document/pdr still readable. See the
retrieval guide for bounded
search, hot/archive lookup, content limits, explicit bytes and refusal handling.
This source contract requires a matching Ledger runtime and a separately reviewed
skills release; it is not a claim that installed flat get already supports all
kinds. Inspect get --help; on task-only versions use native record get if
available, not trial-and-error typed commands. Source merge does not install,
activate, publish, or overwrite customized skills.
Skills
| Skill | Purpose |
|---|---|
ledger-tasks-yylo | Retrieve any Ledger Record by ID and operate tasks, dependencies and source-of-truth boundaries. |
wiki-yylo | Find and maintain durable Markdown knowledge as revisioned Ledger Records. |
workflow-yylo | Store and validate workflow Records while keeping execution separately authorized. |
artifact-yylo | Capture and inspect durable, provenance-bound Ledger evidence. |
benchmark-yylo | Prepare reusable cases and frozen project/task checklists; compare models/harnesses and independently evaluate retained outputs. |
understand-project-yylo | Inspect the user project before planning or implementation. |
plan-ledger-tasks-yylo | Create a PDR and implementation-sized Ledger tasks. |
ralph-loop-yylo | Execute exactly one explicitly assigned Ledger task through validated delivery. |
The benchmark-yylo skill (for YYLO Benchmark / yylo-benchmark requests)
is included starting with version 2.0.4. Current source targets Benchmark 0.2.1:
historical tasks, arbitrary prompts and workflow prefixes; Pi, command and
Workflow Runner adapters; later independent judges without candidate reruns.
Execution, checks, opinions, errors and disqualification remain separate. Study
protocols may require pilot approval, but the runner adds no such ceremony.
Checklist judging guidance and reusable examples are part of this same skill at
skills/benchmark-yylo/references/checklists.md.
There is no separately packaged benchmark-checklist skill and no runtime-owned
skill copy. Benchmark ships its CLI/API documentation and tests; this repository
owns reusable agent instructions. Inspect installed criteria flags before using
checklist features. Source updates here do not publish or activate a skills release.
The four Ledger Record skills are usable with standalone Ledger. The planning and execution skills rely on YYLO orchestration. Each skill has one canonical cross-agent source under skills/<slug>/; agent-specific variants are added only for demonstrated runtime incompatibility.
Ledger Record namespaces must be present in the installed yylo-ledger --help (or delegated yy ledger --help) before an agent uses them. Ledger stores and validates workflow Records but does not execute them.
Repository layout
skills.sh.json
skills/
├── artifact-yylo/
├── benchmark-yylo/
├── ledger-tasks-yylo/
├── plan-ledger-tasks-yylo/
├── ralph-loop-yylo/
├── understand-project-yylo/
├── wiki-yylo/
└── workflow-yylo/
skills.sh.json controls grouping on skills.sh; it does not alter installation behavior.
License
MIT