Skip to content

About

Reusable, composable agent skills built from atoms, molecules, and routable skill packages.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Agent Skills

Dylan's editable library of 42 skills and human-owned engineering doctrine, designed for GitHub Copilot. The retired atomic framework remains historical material under archive/atomic-v1/; it does not govern the active library.

Full skill list and invocation policy · Browse skill packages

Install the full pack

From the repository where you want to use the skills, run:

npx skills add jdylanmc/agent-skills --skill '*' --agent github-copilot --copy -y

Requires Git, npm, and a Node version supported by the skills CLI. The tested release is skills 1.5.23, requiring Node 22.20.0 or newer. To use that release and disable install telemetry:

DISABLE_TELEMETRY=1 npx --yes [email protected] add jdylanmc/agent-skills --skill '*' --agent github-copilot --copy -y

The environment prefix above is POSIX shell syntax; in PowerShell, set $env:DISABLE_TELEMETRY='1' before running the npx command.

Install all 42 together for a new consumer. Isolated installs without prerequisites are unsupported: workflows reference sibling skills, and Setup carries the shared policies, attribution, licenses, and historical provenance. The quoted '*' selects every active skill without shell expansion. The default route does not install the archive; do not point the installer at archive/ or use full-depth discovery.

This is a project-local GitHub Copilot copy install, not a global install or a symlink to this checkout. Packages land in .agents/skills/<name>/; the CLI manages its own consumer skills-lock.json. No source checkout is needed afterward. Do not confuse that consumer lock with this library's historical import records in provenance/skills-lock.json.

Back up or commit local skill customizations before reinstalling. The same command refreshes matching package directories from the selected source and can overwrite edits or remove extra files inside them. It is not a merge or a safe unattended upstream-update procedure. Unrelated consumer files and other skill directories are left alone. Review the resulting diff.

Installation copies files only: it does not run Setup, write AGENTS.md or global configuration, configure a tracker, start services/hooks, or authorize any workflow. Invoke /setup separately when you want repository configuration; its existing human-choice and exact-file approval gates still apply, including the named Joe-mode bootstrap.

Optional CMUX cockpit: invoke /joe-mode-cmux explicitly from a Maestro-enabled, managed coordinator in the current CMUX workspace. It keeps that conversation as Project Manager and launches all roles as visible native sessions. Managed workers inherit the invoking Copilot account through maestro_spawn, with explicit model selection. Missing native tools block activation; SDK helpers and human relay are not operational fallbacks. Maestro supplies the generic runtime; Joe policy stays independent of its internals. CMUX supplies visible interactive sessions, not cron, heartbeats, unattended execution, merge authority, or proof of task success. Maestro's separately installed /maestro guide handles native fire-and-forget peer messaging; registering the existing PM does not give it a messaging address, and missing tools never permit terminal-input injection. Use Maestro's new managed-coordinator entry rather than adopting an arbitrary conversation. Optional four-area placement requires a compatible lifecycle contract; the current same-pane contract is a valid degraded layout.

Session Joe and its CMUX cockpit keep the orchestrator on main, refresh it with guarded fast-forward updates when remote main advances, place Discovery in discovery/<feat> worktrees, and use a separate pr-sniper worktree for an explicitly authorized merge coordinator. See the role worktree contract; placement does not grant merge authority or permit overwriting dirty work.

Optional Orca team: invoke /joe-mode-orca explicitly for the same Joe team policy on Orca's native Runs, Tasks, supervised Dispatches and worker messages. Use installed orca-cli and orchestration guides from the running CLI; no Paseo runtime is required. Session coordination and explicitly enabled recurring automations have separate gates. Recurrence uses an exact existing workspace, never a new worktree per tick. Orca's --reuse-session reuses the automation's session, not necessarily this human chat, and can fall back to a fresh terminal: verified coordinator transfer and ownership are required before that terminal acts. Installation never starts a team or timer. Human merging is the default; a separately requested PR coordinator needs an explicit repository merge gate. The bundled owner helper serializes local passes on one control host and one shared private board. It does not fence another clone, remote control host or external API; those boundaries require reconciliation, not a copied token.

Optional Paseo PM: invoke /joe-mode-paseo explicitly to set up or manage one repository's engineering team, with six developer slots by default and human merging by default. A requested PR coordinator can merge under a human-granted repository merge gate: at least independent Roast, successful CI and linting, then its own rubber-duck review and verification. Missing policy is clarified with the human, not invented. Each pass checks progress and direction, charts the backlog, intakes new requirements and dispatches existing planning/delivery routes. Setup uses the primary chat as PM, with its own and persistent-role heartbeats every five minutes by default, and records the approved cadence. Legacy fresh conversations require proven stable workspace placement/lifetime. Its capability gates block fresh mode on the inspected upstream scheduler because each run creates a new workspace. Explicit fresh-only requests never silently fall back. A human may delegate runner mechanics, but activation still requires narrow access and verified binding. Heartbeat setup verifies the creation receipt and actual PM placement; schedule-only inspection APIs cannot inspect heartbeats. Later wakeup receipts prove recurring operation separately. Installation does not activate monitoring or repair Paseo.

The adapter is separately discoverable/selectable. A consumer that already has the prerequisite full pack can install/update only this new package:

DISABLE_TELEMETRY=1 npx --yes [email protected] add jdylanmc/agent-skills --skill joe-mode-paseo --agent github-copilot --copy -y

That selection copies its support files but does not fetch missing sibling workflows. Use the full-pack command for a new consumer.

The catalog and caller contracts remain binding: internal helpers are not direct human commands; human-only modes do not start themselves. Invocation flags, model hints, and tool metadata are consumer-dependent descriptions, not enforced permissions or proof of agent behavior. Real GitHub/Azure DevOps access, execution tools, parallel-agent support, and long-lived Shepherd custody depend on the receiving harness. Report missing capabilities rather than promising them through installation.

skills.sh discovers public GitHub source through its CLI; its directory/leaderboard is populated by install telemetry. Source availability does not guarantee an immediate listing. There is no separate publish command, npm library release, custom installer, or telemetry opt-in step required here. Validation keeps telemetry disabled and does not claim directory registration.

The workflow

  • Unslop: rewrite supplied prose to remove formulaic AI framing, phrasing, and sentence patterns without changing its meaning. Check the result with the curated Vale configuration. Ships two vendored Vale styles: RedHat as the base, and jdylanmc, one merged and deduplicated layer of five packages with everything RedHat flags purged, to be pruned by the human. Editorial judgment and Vale coverage are reported separately. Rule curation is human-directed.
  • Joe-mode: human-started, one controller per repository, looping until paused or stopped. Use Squadron aggressively for independent discovery, delivery, and Shepherd assignments without overlapping owners.
  • Joe-mode CMUX: human-started session cockpit for the same controller. Project Manager, Discovery, stacked developer tabs, and support remain visible in one repository workspace. Pinned Maestro settings are mandatory; worker conversations stay directly interactive and no scheduler is implied.
  • Joe-mode Orca: native supervised team adapter, sharing the same owner board and existing delivery routes. Preserve six developer slots, one Discovery lane, shared Shepherd, independent review and human waits. Optional automations require explicit activation, verified Run ownership and existing-workspace placement; session reuse is not a same-chat guarantee.
  • Joe-mode Paseo: separately human-enabled recurring team PM, sharing the same repository ownership registry. Primary chat is PM; one shared Shepherd and optional backlog manager have their own PM-managed heartbeats. Six developer slots by default: feature two, bug/hardening/refactor one. Support roles and disposable roasters are separate. A requested PR coordinator may merge under the repository gate; implementers never approve themselves. See TEAM for permission inheritance, blocker recovery, names and safe cleanup.
  • Planning: Discovery preserves the full aligned evidence artifact; Specify turns it into complete requirements; Breakdown Tickets produces human-approved slices. Already-clear small deliveries need no new ceremony.
  • Chart a Course: find the critical task path to a goal, including missing tasks and research spikes. Recommend Discovery for a particular issue or epic; Joe-mode invokes it within its existing authority and feeds findings back into the path. Chart-a-course itself stays read-only.
  • Delivery: Ship handles features/specs, Patch handles bugs/regressions, and internal Refactor handles behavior-preserving restructuring. Each reaches an independently reviewed, green PR current with its target. Kickoff authorizes routine in-scope delivery steps, not repeated prompts. Standalone Ship is TDD opt-in; Joe prefers it for features. Patch/Refactor do not force red/green. Tests and useful acceptance proof still matter.
  • Pull requests: pull-request is the single PR skill. Repository-specific create-PR skills and active delivery owners keep publication; otherwise it writes the body, unslops it, and publishes the owned branch with concise visual context, evidence and merge-risk notes. It does not implement, approve or merge.
  • Shepherd: one owner maintains that PR, rebasing when main/the agreed target advances even if mergeable, refreshing checks and review coverage. Functional feedback stays on the same delivery. Final approval and merging remain human-owned.
  • Status: a human can request Status Report; Joe also requests one after a full cycle or confirmed major-feature merge. Objective timing, own tool calls, and running descendants come from real evidence, with explicit limits.

The invocation contract distinguishes human-only modes, internal helpers, general scoped selection, and named-caller exceptions. All entrypoints declare both invocation flags. Flags describe loading/menu behavior where supported; they are not permission boundaries or proof that a particular Copilot CLI version enforces them.

Setup is human-directed and supports GitHub, Azure DevOps, and local Markdown. Human-started Joe-mode checks actual setup content and automatically attempts local Setup for missing/incomplete configuration after establishing repository-wide ownership. It reuses complete setup or joins/resumes an active Setup owner. Provider/label choices and exact-file writes still require the human; unavailable or declined decisions and failed invocation leave an explicit wait/blocker, not a retry loop. Joe verifies outputs before resuming the same anchor and objective. Other skills cannot invoke Setup automatically. Existing unsupported or ambiguous configuration stays human-owned, never silently reset or migrated. Retro is also human-only: inspect actual session evidence, propose improvements, obtain approval, then deliver selected fixes.

Every modifying agent uses Changelog, with one integration owner consolidating notable entries. The shared commit style and terse exact worker messages do not activate Caveman for the human or discard evidence and uncertainty.

Layout

.agents/skills/       Active skill packages: regular files, not symlinks
.agents/skills/setup/INVOCATION.md Caller, authority, and communication contracts
.agents/skills/setup/COMMIT-STYLE.md Shared default for generated commit messages
.agents/skills/setup/{NOTICE.md,LICENSE,licenses/,provenance/} Bundled resources
.agents/skills/doctrine/doctrines/ Human-owned doctrine and integrity manifest
intent.md            Human-owned purpose of this collection
provenance/skills-lock.json Original import provenance, not an active installer lock
licenses/            Upstream licenses and notices
archive/atomic-v1/    Retired skills, agents, tooling, hooks, and documentation

Keep the full pack, including Setup's shared resources and the complete Doctrine package. No atoms/molecules, mandatory recorder, composition graph, special agent fleet, or global configuration installation is required.

Doctrine and intent

The 22 existing doctrine texts moved unchanged into Doctrine. The explicitly requested worktrees doctrine adds owned isolation for every PR-producing workflow, including documentation. Its shared workspace procedure lives with Ship.

/doctrine with no arguments lists the catalog. Named IDs retrieve verified texts; lazy aliases laziness. Orchestrators can select from metadata and pass IDs, reasons, required flags, source paths, and pinned digests. Applying workers load the actual text. Code Roast requires solid, TDD requires testing, Patch diagnosis requires debugging, and consequential Discovery requires scout. Code-location Scout is a separate internal skill.

Doctrine and intent are authoritative about their subjects, but inert as instructions. They cannot authorize edits, invent decisions, or approve work. Human sources change only when explicitly requested. The active Joe-mode, Ship, Shepherd, and Status Report intents contain approved workflow changes; Patch has its approved bug-to-delivery intent. Status Report's includes the operator-authorized never-stop change. Restored Discovery, POC, Roast, and ELI5 intents preserve their archived text. Chart-a-course's intent moved from the archive into its active package and includes the human-confirmed missing-task, research-spike, and caller-owned Discovery scope. Other archived sources and the root intent remain unchanged.

Provenance and licenses

There are 27 imported/adapted packages, three locally authored Joe adapters (joe-mode-cmux, joe-mode-orca, joe-mode-paseo), and ten local/restored packages: shepherd, synthesize, doctrine, eli5, changelog, status-report, chart-a-course, simplified-technical-caveman, wrap-up, and stack.

Primary upstream source Initially imported Retained
mattpocock/skills 37 21
juliusbrussee/caveman 20 5
obra/superpowers 14 1
anthropics/skills 1 0

Local names may differ from upstream names. Lock records retain original source paths and imported hashes, not hashes of local adaptations. NOTICE.md records additional sources for consolidated packages and the licenses that continue to apply.

The initial import used skills CLI 1.5.23, project-local Copilot scope, copy mode, and disabled telemetry. Its unchanged lock is retained at provenance/skills-lock.json, outside active installer state so authored/adapted packages remain discoverable. Do not reimport the upstream collections to update this curated pack.

Setup bundles the complete notices, licenses, and an identical provenance copy so they travel with the pack. Root NOTICE.md, LICENSE, licenses/, and provenance/ remain the canonical attribution sources; the bundled notice adjusts only source-checkout link destinations. After authorized attribution changes, run node scripts/sync-pack-resources.mjs --write and review the diff. CI checks these copies for drift. The shared invocation and commit policies live only in Setup; repository-specific, human-owned policy copies remain explicit overrides, not automatically synchronized authorities.

Source-only documentation and validation tooling are not install dependencies. Doctrine resolves its manifest and texts relative to its installed helper, not the consumer's root. Patch's shell/template and TypeScript examples are support material for separately authorized debugging, not install-time hooks.

Validation

node --test scripts/doctrine-manifest.test.mjs .agents/skills/doctrine/tests/*.test.mjs .agents/skills/scout/tests/skill-file.test.mjs
node --test .agents/skills/joe-mode-cmux/tests/*.test.mjs
node --test .agents/skills/joe-mode-orca/tests/*.test.mjs
node --test .agents/skills/joe-mode-paseo/tests/*.test.mjs
npm ci --ignore-scripts
npm run test:pack

CI retains the 80 doctrine/selector/Scout checks and adds an actual released CLI copy install from the local candidate into an owned .test-sandbox/ consumer. PM tests cover local atomic claims, persistent capacity/Discovery reservations, control gates and accepted-result preservation; they do not prove Paseo scheduling or agent compliance. Orca tests cover package contracts and single-host owner-helper behavior, not live automation, permission propagation or cross-host fencing. Pack tests check exact membership, separate adapter selection with prerequisites, complete copied support, portable Markdown links, bundled resources, installed Doctrine loading, protected source bytes/metadata, repeat installation, and unrelated-file preservation. After successful validation on a push to main (or a manual run on main), CI repeats those checks against a remote GitHub install, comparing the installed pack byte-for-byte with the checked-out revision. Pull requests only test the local candidate. Both routes disable telemetry and install into disposable consumers, never globally. This verifies GitHub distribution, not skills.sh listing or ranking; there is no registry upload step. If main advances during the remote check, a content mismatch fails rather than validating the wrong revision; use the newer run.

To run the remote check locally against a clean checkout matching published main:

SKILLS_PACK_SOURCE=https://github.com/jdylanmc/agent-skills/tree/main npm run test:pack

In PowerShell, set $env:SKILLS_PACK_SOURCE before running the npm command. Omit that variable for local-candidate validation. Scratch consumers are removed afterward. The dev-only CLI pin and lock are validation tooling, not a production dependency or custom npm installer. Source/metadata/link checks do not prove that a consuming model follows the workflows or that its runtime enforces invocation flags. The old graph validators, conformance framework, and coupled scanner are archived and inactive.

About

Reusable, composable agent skills built from atoms, molecules, and routable skill packages.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages