An opinionated engineering standard packaged as a Claude Code plugin: spec-driven development, a team's good practices, and per-repository context — with automatic onboarding and, above all, a development flow that is enforced by hooks instead of described in prose.
Status:
0.3.0— 25 skills, 76 playbooks. Covers the full cycle from idea to merge for PM, QA and dev. Everything team-specific comes from your team config: runsetuponce and the team is configured; runonboardin a real repository and it fills the context from the code.
📚 Playbooks — the "how" of each skill lives in
references/, loaded on demand. Full index: PLAYBOOKS.md (auto-generated).
Most agent setups describe a process and hope the model follows it. It doesn't — not reliably. An agent under pressure skips the boring step, reorders the flow, or reports a step as done when it wasn't.
So the process here is mechanical, not textual:
- A hook blocks (
exit 2) editing code before an issue exists and is assumed. - A hook blocks opening a PR before local review is closed with zero findings.
- A hook blocks moving an issue to review before the PR is open and the cost is recorded.
- Milestones are stamped from what the tool actually did, never from the agent's claim.
- State lives in the repository (
.git/ferion-flow), so a new session resumes instead of restarting. - The only way around it is an explicit
bypass "<reason>"— recorded, and declared in the summary.
Everything else in this plugin — the constitution, the 25 skills, the playbooks — is what makes each of those steps worth taking.
- Claude Code
ghauthenticated (gh auth status) — this is the only hard dependency. Issues, PRs and labels all go through it.- Python 3 (used by the cost script and the validation script).
This repository is a Claude Code marketplace:
/plugin marketplace add FerionAI/ferion-engineering
/plugin install ferion-engineering@ferionReopen the session to load it. Confirm with /plugin list (ferion-engineering@ferion · enabled) —
the session-start hook appears and the skills become available without a prefix (e.g. /start,
/ship, /preflight). To update:
/plugin marketplace update ferion && /plugin update ferion-engineeringNot a developer (PM/QA)? Ask someone to install it once; after that, just talk to it in plain language — start with
/setup(configures) or/start(does the work).
Every code change goes through 6 steps, in this order, with no skipping:
| # | Step | Closes when… | Where |
|---|---|---|---|
| 1 | issue | the issue exists — the person cited it, or you asked and created it | issues |
| 2 | in progress | you assumed the issue (assignee) and labelled it status:in-progress |
issues |
| 3 | implement | spec/plan in specs/<feature>/ and the code in a vertical slice with tests |
spec-flow → ship/fullstack |
| 4 | local review | review + preflight looping until zero findings |
review, preflight |
| 5 | PR + cost | PR open with Closes #N and tokens/USD/hours recorded on the issue |
integrations, cost |
| 6 | in review | issue labelled status:in-review, and said so in the summary |
issues |
Check where you are at any time:
bash hooks/flow-gate.sh status| Component | Type | What it is for |
|---|---|---|
setup |
Skill | Guided configuration, no code — connects tooling and collects the standard through simple questions (designed for PMs and QAs too). |
start |
Skill | Natural-language entry point — understands the request in plain language and routes it, without anyone learning skill names. |
ship |
Skill | Idea-to-PR combo — orchestrates spec → vertical slice → integrations → quality → review → PR. |
spec-flow |
Skill | Spec-driven flow (specify → clarify → plan → tasks → implement → analyze) anchored in the standard. |
issues |
Skill | GitHub Issues as the source of truth: read from the issue, write back to it, Closes #N closes the loop. |
epic |
Skill | For PMs: turns an epic into well-formed issues with stack, testable acceptance criteria and the Definition of Done. |
discovery |
Skill | Design thinking around the pipeline — empathize (analytics → brief), ideate (diverge/converge), test (usability + post-ship behavioral validation). |
preflight |
Skill | Pre-PR gate — reviews everything against every rule + runs the real tools + independent anti-hallucination verification, looping until zero automated findings. |
review |
Skill | Human review after preflight: code taste and "does anything look off?". |
qa |
Skill | QA kit: test cases from the issue, bug reports, risk/regression, exploratory testing, automation, UX/a11y QA, production verification and the approval gate. |
quality |
Skill | Quality and security rule base: OWASP, Nielsen, WCAG, Core Web Vitals, static analysis and DORA metrics. |
design-review |
Skill | Pre-code design gate — validates a layout/Figma against the Design Constitution before building; a UX linter with severities. |
design-system |
Skill | Discovers the project's design system and drives its progressive adoption — new UI uses it, legacy migrates without blocking deploy. |
privacy |
Skill | Personal data done right (GDPR/LGPD/CCPA): data map, legal basis, retention, masking in logs and analytics, data subject rights. |
constitution |
Skill | The non-negotiable standards (architecture, tests, review, typing, observability) + Definition of Done. |
context |
Skill | Application inventory, infrastructure patterns and per-language conventions — the context that stops everyone from starting from zero. |
onboard |
Skill | Initializes the standard in a new project (scaffold) or an existing one (scans, documents, reports gaps). |
agentic-flow |
Skill | How to work with AI efficiently: less code, response optimization, model selection per task, subagent orchestration. |
cost |
Skill | Measures cost per task (tokens + time) and records it on the issue; feeds efficiency tracking and feature pricing. |
fullstack |
Skill | Strengthens front-end people in the back end and vice versa (mentor mode + guardrails), enabling vertical slices. |
workspace |
Skill | Multi-repo coordination — understands how the repos of a workspace relate (polyrepo, front ↔ N backs through a contract) and orchestrates the slice, the PRs and the deploy order. |
integrations |
Skill | External tooling expertise: decoupled, observable integration code + MCP-first operation. |
health |
Skill | Living engineering health dashboard (DORA + quality gate + observability + Core Web Vitals). |
release |
Skill | Shipping to production safely: deploy strategy, feature flags, rollback criteria and mechanics, SemVer from the commits, zero-downtime data migration. |
incident |
Skill | Incident response + blameless postmortem: severity, mitigation, communication, root cause and tracked action items. |
| Always-on hook | Hook | Injects the standard at session start and on every code change, and enforces the flow order. Also blocks AI signatures on commits and PRs (P3). |
.mcp.json |
MCP | Optional MCP connections (GitHub, Playwright). |
Call the setup skill: it walks anyone — including people who do not code — through connecting
the tooling and configuring the standard, ending with everything ready. After that, a PM can say
"break down this epic" and "deliver this issue", and the plugin takes it to a finished PR; the
engineer only does the final review.
The entry point is start: describe what you need in plain language — "build this feature",
"is this secure?", "onboard me into this repo", "how are our metrics", "can this leak personal
data?" — and it routes you. For a full end-to-end feature, the ship combo orchestrates
spec → vertical slice → integrations → quality → review → PR.
AGENTS.md + memory/constitution.md are the portable core: drop them into a repository and
Cursor, Gemini CLI or Codex get the principles, the Definition of Ready/Done and the spec-driven
process. Skills are a Claude Code feature; the standard is not. See
docs/multi-llm-distribution.md.
Nothing here is hardcoded to one company. Anything team-specific is either a <FILL: ...>
placeholder or lives in config/team-config.md (git-ignored), filled by setup and onboard.
CI enforces that with scripts/check-leaks.sh.
Fork it, change the constitution, keep the flow. That is the point.
MIT — see LICENSE.