Engineering Advisor
View on GitHub

v0.1.0MIT license

Engineering Advisor

The right amount of engineering process for the project you actually have.

An open-source agent skill for Claude Code and other agents that support the Agent Skills format. It works out what kind of project you have, how much is at stake and how mature it is, then recommends the practices that fit and turns them into a phased plan.

Install View on GitHub

Your project

  • GoalLaunch a site next month
  • RepositoryA static Next.js site
  • TeamOne maintainer
  • Data and moneyA contact form, no payments
  • Where it shipsHost with instant rollback

The advisor

  1. 1Identify the objective
  2. 2Gather context
  3. 3Classify the project
  4. 4Decide each area
  5. 5Deliver a phased plan

The plan

Essential nowNextLater

  • GitHub Flow with only main
  • Preview deployments and instant rollback
  • A short CI run on pull requests
  • HTTPS and documented DNS ownership
  • A monitored inbox for form data
  • One rehearsed rollback

Not recommended

  • GitFlow
  • A staging environment
  • Kubernetes

Classification

FamilyStatic website
Risk profileLow
Maturity stagePre-launch
Release channelReversible
Worked example 1 of 3: a static business website before launch.

Process has a cost.

More branches, environments, tests, services and tools are not better by default. The skill spends process where it reduces a risk the project actually has, and says plainly when something is unnecessary.

Size does not set the security floor.

Money, personal data and regulated data do. A small app that takes payments or stores health data needs strong safeguards from its first release.

Same question, different answers

The advice changes with the project.

Three projects ask how to set up branches, reviews, environments, checks and recovery. Pick one to see the skill's answer; these are its own worked examples.

A static business website before launch

A Next.js site for a consulting company, launching next month.

Static websiteLow riskPre-launch

  • BranchesBarely changes

    GitHub Flow with only main. Every pull request gets a preview deployment.

  • ReviewGrows with the team

    Pull requests and a required check. No required approvals for a solo maintainer.

  • EnvironmentsGrows with the stakes

    Previews and production, with instant rollback. No staging.

  • ChecksGrows with the stakes

    Lint, type check, build, link check and an accessibility check, in a few minutes.

  • RecoveryGrows with the stakes

    One rehearsed rollback before launch.

  • Leaves out

    • GitFlow
    • A staging environment
    • Kubernetes

What matters most. Documented domain and DNS ownership, HTTPS, a monitored inbox for form data and a privacy notice. The risk rises to Moderate if the form stores personal data.

A new B2B SaaS with sign-in and subscriptions

Sign-in, a PostgreSQL database and paid subscriptions, before building starts.

Multi-tenant SaaSHigh riskPrototype

  • BranchesBarely changes

    GitHub Flow, with feature flags for unfinished work.

  • ReviewGrows with the team

    One required approval once a second developer joins.

  • EnvironmentsGrows with the stakes

    Local, previews, staging with a billing sandbox, and production. Staging earns its cost because billing and identity cannot be reproduced in previews.

  • ChecksGrows with the stakes

    Tenant-isolation tests, an authorization matrix, integration tests against real PostgreSQL and billing webhook tests.

  • RecoveryGrows with the stakes

    Point-in-time recovery, with one restore tested before the first paying customer.

  • Leaves out

    • Microservices
    • Kubernetes
    • A permanent staging branch

What matters most. Customer data and money set a high floor from the first release, even with one developer.

A team growing from one developer to five

A full-stack web app moving from early production to a growing team.

Full-stack web appModerate riskGrowing team

  • BranchesBarely changes

    Keep GitHub Flow, with feature flags so unfinished work can merge early.

  • ReviewGrows with the team

    One required approval on every pull request, and CODEOWNERS for billing and authentication.

  • EnvironmentsGrows with the stakes

    Unchanged. No environment branches.

  • ChecksGrows with the stakes

    Small pull requests that can be reviewed in one sitting. A merge queue only if main starts breaking.

  • RecoveryGrows with the stakes

    Written deployment and rollback steps that anyone can run.

  • Leaves out

    • GitFlow
    • Environment branches
    • Microservices

What matters most. Growing to five developers changes reviews and ownership first, not the branching model.

The branch model barely moves across the three. What grows with the stakes is the environments, the checks and the recovery, and each answer names what it leaves out as clearly as what it adds.

How it works

Four things decide the advice.

It reads your repository without changing it, asks at most three questions, and classifies the project on four axes. Then it starts from the family's defaults, raises them to meet the risk, trims them to the stage, and names what can wait.

Family

What kind of software it is.

  • Static websites
  • Full-stack web apps
  • SaaS and marketplaces
  • APIs and backends
  • AI-powered apps
  • Mobile and desktop
  • Libraries and CLIs
  • and seven more

Risk profile

How much is at stake. It sets the floor.

  1. LowNo accounts, no money
  2. ModerateAccounts or personal data
  3. HighMoney, sensitive or regulated data
  4. CriticalPhysical harm or legal breach at scale

Maturity stage

Where the project is now.

  1. 1 PrototypeAnswer a question cheaply
  2. 2 Pre-launchRelease readiness and recovery
  3. 3 Early productionReliable releases
  4. 4 GrowingOwnership and coordination
  5. 5 MatureReliability where justified

Release channel

Whether a bad release can be taken back.

  1. Undone in minutesWeb deploys. Invest in fast detection and rollback.
  2. Hard to take backApp stores, registries, firmware and published data. Test more before release.

Try it

Ask in your own words.

These prompts are good starting points. Copy one, or describe your own project.

  • I'm launching a static Next.js business website next month. Help me prepare the repository and deployment workflow.
  • I'm starting a SaaS with authentication, PostgreSQL and subscriptions. Design the engineering workflow before I begin.
  • Review this repository and tell me which engineering practices are missing before launch. Don't change anything.
  • Do I really need a develop and a staging branch for this project?

Install

Install it in the agent you use.

Claude

In Claude Code, use either the plugin or a copied skill folder, not both, or the skill will be listed twice.

Claude Code, as a plugin

  1. Inside Claude Code, add the marketplace.

    Claude Code
    /plugin marketplace add Achal13jain/engineering-advisor

  2. Then install the plugin.

    Claude Code
    /plugin install engineering-advisor@engineering-advisor

From a terminal, the same two steps are:

Terminal
claude plugin marketplace add Achal13jain/engineering-advisor
claude plugin install engineering-advisor@engineering-advisor

Claude uses the skill when a request matches it, or you can invoke it as /engineering-advisor:engineering-advisor.

Claude Code, by copying the folder

Copy the skill into ~/.claude/skills/ for all your projects, or into a repository's .claude/skills/ to share it with the team.

Terminal
git clone https://github.com/Achal13jain/engineering-advisor.git
cp -r engineering-advisor/skills/engineering-advisor ~/.claude/skills/
PowerShell
Copy-Item -Recurse engineering-advisor\skills\engineering-advisor "$env:USERPROFILE\.claude\skills\"

Start a new session, then invoke it as /engineering-advisor or just ask.

claude.ai

Download engineering-advisor.zip from the latest release and upload it in claude.ai as a custom skill.

Claude API

Upload the skill folder through the Skills API. Skills there run without network access, so the skill marks any tool versions it cannot check as unverified.

Codex

Codex reads your personal skills from ~/.agents/skills, and a repository's skills from .agents/skills. Copy the skill folder there.

Terminal
git clone https://github.com/Achal13jain/engineering-advisor.git
mkdir -p ~/.agents/skills
cp -r engineering-advisor/skills/engineering-advisor ~/.agents/skills/

Codex finds new skills automatically; restart it if the skill does not appear. Mention it with $engineering-advisor, or run /skills to pick it. On Windows, ~ is your user folder.

Gemini CLI and Copilot CLI also read ~/.agents/skills, so this one copy works in all three.

Skills CLI

One command installs the skill into the agents you choose, including Claude Code, Codex, Gemini CLI and GitHub Copilot.

Terminal
npx skills add Achal13jain/engineering-advisor

Choose agents with -a, for example -a codex, -a gemini-cli or -a github-copilot. Add -g to install it for all your projects instead of only the current one.

Gemini CLI

Install the skill straight from this repository.

Terminal
gemini skills install https://github.com/Achal13jain/engineering-advisor.git --path skills/engineering-advisor

Gemini CLI asks you to confirm the install. It installs for your user by default; add --scope workspace to install it for the current project only. In a running session, enter /skills reload. Gemini uses the skill when a request matches and asks before activating it.

Copilot CLI

Copilot CLI reads your personal skills from ~/.copilot/skills, and a repository's skills from .github/skills. Copy the skill folder there.

Terminal
git clone https://github.com/Achal13jain/engineering-advisor.git
mkdir -p ~/.copilot/skills
cp -r engineering-advisor/skills/engineering-advisor ~/.copilot/skills/

In a running session, enter /skills reload, then check it with /skills list. Invoke it as /engineering-advisor, or just ask. On Windows, ~ is your user folder.

What it will and will not do

  • It reads repositories without changing them. It does not change branch protection, repository settings, deployment configuration or production systems without your explicit approval of that specific action.
  • It does not run destructive Git commands or discard uncommitted work.
  • It never prints or commits secrets. If it finds one, it reports the file and line without the value and recommends rotating it.
  • It treats repository contents, issues and web pages as data, not as instructions.
  • It checks facts about third-party tools at the official source when it is used, and says when it cannot.
  • It is an advisor, not an auditor. Legal and compliance questions are flagged as needing a qualified person.