Shippable Skills

Name: shippable-skills
Author: renzrollon

Agent skills other engineers can install on Cursor, Copilot, Codex, or Claude Code, gated before they ship.

npx skills add renzrollon/shippable-skills

Then type /session-retro to score this conversation, /product-spine to act as product owner for your repo, /bump-release to bump every manifest version at once, or /session-handoff to save where a session stands before a reset. fix-tests, explore, and session-primer are model-invoked: the agent reaches for them when the task fits.

The lead skill retros the live session. session-retro scores the conversation you are still in and folds repeatable steps back into the skills you called.
Install the folder, keep the files. You keep ordinary skill files in your project or home directory, so the next lesson stays. Updates later: npx skills update.
Gated before they ship. The release pipeline (SkillGantry) lints the skill, runs its evals, scans it, and proves the archive installs. The stages live in CONTRIBUTING.md.

Install

The npx path needs an Agent Skills host and npx. Claude Code may use the plugin instead.

npx skills add renzrollon/shippable-skills

Then type /session-retro in any session. First hour: session-retro.

Claude Code can take the same set as a plugin instead. Do not also run the npx installer, or every skill lands twice:

/plugin marketplace add renzrollon/shippable-skills
/plugin install shippable-skills@shippable-skills

How it works

The host copies the skill folder. You invoke the skill. The model follows SKILL.md.

You type	What happens
`npx skills add renzrollon/shippable-skills`	Skill files land in the project or home directory you chose
`/session-retro`	The host loads session-retro and runs the retro on this conversation
`/session-retro --focus tokens`	Same snapshot, token section only
`/session-retro --focus wiring`	Same snapshot, wire-in section only
`/product-spine`	The host loads product-spine and proposes work threads from the repo
`/bump-release`	The host loads bump-release, finds every manifest that carries the version, and waits for your yes before it edits, commits, tags, or pushes
`/session-handoff`	The host loads session-handoff and writes where this session stands to a handoff file

You type /session-retro, /product-spine, /bump-release, or /session-handoff. The agent will not start any of them in the middle of your work.

The model-invoked skills have no command to type. fix-tests loads when tests are failing, explore when a question spans several subsystems, and session-primer at session start or on "where was I". You can still ask for any of them by name.

Why product-spine

Most PM skill libraries teach a method in chat: a story template, a RICE table, a PRD outline. The artefact, if any, does not live next to the code, and nothing checks that a brief still ladders to a strategy bet.

product-spine is repo-native. Bare invocation is an engineer-as-PO coach: it fans out over the repo and proposes Feature / Improvement / Fix threads on a mermaid map. Picking a thread lands it on three files at fixed paths. It refuses to invent the platform.

Threads before templates. What to work on next, grounded in evidence (or marked thin), not a blank form. Draft is how a picked thread lands.
A greppable ladder from brief to epic to one-pager initiative, so "every brief serving this bet" and "every brief if we kill this bet" are grep lines, not a hunt through chat.
No PRD and no user stories. The brief's four sections (need, platform pieces, constraints, reasons) are the unit; discovery sits in the epic's sibling folder. Spec and how live in the change, not here.
Primer-grounded capability. "Can users do X today" cites docs/product/context/platform-primer.md, or answers Cannot confirm if that file is missing.
Decisions inline. Trade-offs land in the brief, the epic, or the one-pager change log. There is no decisions folder.

New files stay Status: draft until you flip them. Docs: product-spine.

Why this

A folder you own, not a prompt you paste. The installer writes files you can open. See session-retro.
A gate, then a release. Lint, evals, security, then an archive that has been installed once. See CONTRIBUTING.md.
A retro with no gather step. session-retro scores the session it is running inside and refuses an on-disk transcript. See the docs.
Wire-ins that name the file that runs. The lesson goes into the skill you called, not a new SKILL.md drafted on the spot. See Where wire-ins land.
Host-neutral by construction. Cursor, Copilot, Codex, Claude Code, and the rest of the skills.sh hosts. The Claude plugin is an alternative to npx, not a second install.

How it compares

The axis is installable-and-gated versus copy a SKILL.md by hand.

kevinlin/skills is the public set the same SkillGantry pipeline was built around. It ships a delivery-lifecycle catalog as a Claude Code plugin (/plugin marketplace add kevinlin/skills). For a single skill it says to copy the folder.

This repo is the small set you add with npx skills add renzrollon/shippable-skills. The lead skill is a live-session retro, not a spec or review workflow.

The skills CLI is the installer, not a rival collection. skill-lint, skill-up, and SkillSpector are the gate tools, not products you install instead of this.

Status

session-retro, product-spine, bump-release, fix-tests, explore, session-handoff, and session-primer are the promoted skills. Beta work stays in skills/in-progress/ and is not in the plugin. The package on npm is not published; install from GitHub via npx skills add.

Dated gate results for session-retro 0.1.0 (21 Aug 2026) are in its changelog.

product-spine 0.3.0 has not been through this gate yet. That run is the next release step. bump-release, fix-tests, explore, session-handoff, and session-primer (all 0.1.0) have not been through it either.

Conventions

Skills that share state agree on a few paths and shapes. Each skill restates the rule it uses inside its own SKILL.md, because a skill cannot read another skill's files. This section is where they are documented once.

Instruction file. The first existing of AGENTS.md, CLAUDE.md, .github/copilot-instructions.md at the repo root. A skill may write inside a fenced managed block it owns without asking, and names that write in its report. Any edit outside such a block waits for your yes.
Path overrides. Put shippable-state-dir: <path> or shippable-handoff-dir: <path> on a line in the instruction file, with the path relative to the repo root. A --state-dir <path> or --handoff-dir <path> argument, or a directory you name in the request, wins over the file. A path from the file must stay inside the repo: an absolute path or one containing .. is ignored.
State directory. Your path, else the instruction-file override, else .claude/ when it exists, else .agents/.
Handoff directory. Your path, else the instruction-file override, else the first existing of .claude/handoff/, .agents/handoff/, docs/handoff/. Readers scan every existing one. When none exists, writers create .agents/handoff/ and advise a gitignore line. If the write is denied, the writer does not retry and puts the content in its reply. Files are named <kind>-<slug>-<YYYYMMDD-HHMMSS>.md, where kind is handoff, explore, grill, wire-in, or skill-candidate. grill is reserved for a later skill.
Check profile. How a project runs its tests, typecheck, lint, and build lives in a managed block fenced  and  in the instruction file, created in AGENTS.md when none exists. The block is authoritative. An optional cache at <state-dir>/testing/profile.json mirrors it, and when the two disagree the block wins.
Brief files. An explore brief carries the headings Meta, Problem / Intent, Codebase Findings, Critical Files, Options Considered, Recommended Direction, Assumptions Made, Pending Clarifications, Risks / Gaps. A handoff document carries the sections named in session-handoff. A skill that reads either picks the newest matching file, takes what it records as settled, and re-checks anything the repo may have changed since.

Skills

These split on one axis: who can invoke them. User-invoked skills are reachable only when you type them. Model-invoked skills can be invoked by you or reached for automatically when the task fits.

Engineering

User-invoked

bump-release: Bump the project's version in every manifest at once, then commit, tag, push, and cut a release, so no manifest keeps a stale version.

Model-invoked

fix-tests: Discover and persist how a project's tests run, then run the suite and repair failures by root cause: cluster failures by error signature, fix the shared cause once, and assert no test was weakened.
explore: Deep read-only reconnaissance before a change is designed. Fans out parallel investigators across subsystems, then synthesizes one picture and, in an autonomous run, writes a durable brief.

Productivity