Internal AI-Agent Plugin for a Software Studio
An internal initiative at a software delivery studio, built for use across Django, Magento 2, WordPress, and Next.js projects
Context
About this project
Engineering teams at the studio were relying on ad hoc prompting to get consistent results from AI coding assistants, every developer re-explaining the same delivery conventions, review standards, and SEO rules from scratch. I designed and built an internal Claude Code plugin that packages this knowledge into reusable, versioned tooling, spanning requirements, code review, QA, release, and a new structured-data / on-page SEO auditing capability.
Challenge
The challenge
- Delivery conventions (requirements, scaffolding, review, QA, release) lived as tribal knowledge, not as something the AI assistant could apply consistently
- Structured-data and SEO guidance was a prose checklist the model had to re-apply by hand each time, with no way to catch a regression
- No deterministic way to score a page's on-page SEO the same way twice, which makes an audit worthless for client-facing reporting
- Existing checks only read generic robots directives, missing Google-specific overrides that cause real 'why did we vanish from Google' cases
- No repeatable review discipline for AI-assisted changes across the team
Approach
What I built
- 1
Delivery agents and shared skills
Built a set of role-scoped agents, project manager, tech lead, code reviewer, QA engineer, release manager, that orchestrate a shared library of skills across the studio's stacks (Django, Magento 2, WordPress, Next.js, React), so the same conventions apply no matter who's driving the assistant.
- 2
Deterministic scoring engine
Replaced a structured-data checklist the model used to self-apply in prose with Python validators that return a real exit status: syntax, required properties, breadcrumb rules, schema eligibility, staging-URL leaks.
- 3
Cross-skill enforcement
Wired a structured-data checkpoint into both the review chain (code review, QA, release, deploy) and the prevention side (planning and page-scaffolding skills), so issues get caught on the way in and on the way out, with the rules defined once.
- 4
Full on-page SEO auditor
Broadened a schema-only audit idea into a five-category on-page scorer, structured data, core tags, indexability, social cards, media/speed, plus a read-only audit agent that runs in parallel across many URLs, backed by a fixture-based automated test suite.
- 5
Review-driven correctness passes
Ran the system through a real human-plus-automated review loop and fixed what it surfaced: a false negative where Google-specific noindex directives were invisible to the indexability check, and a page-capture bug where an embedded iframe could overwrite the real page's headers. Also made the final score assembly reproducible byte-for-byte across runs, not just the sub-scores.
Results
Outcome
- A reusable plugin, dozens of skills plus several delivery agents, now backs day-to-day requirements, review, QA, and release work
- An SEO and structured-data auditor scores pages 0-100 with the same result on every run, end to end, not just at the sub-score level
- Two real correctness bugs caught and fixed before shipping: a Google-specific noindex false negative, and an iframe page-capture defect
- Deduction values, eligibility rules, and schema logic each live in exactly one place, the auditor delegates to the same rules the writer skill uses
- A rounded 100+ automated tests back the scoring engine so a page can't silently start scoring differently between releases
Key takeaways
- Determinism is a product property, not an implementation detail, a tool that scores the same page differently on two runs is worthless for client reporting
- The 'happy path' hides real correctness bugs, both defects here were caught in review, not in the first implementation
- AI-assisted delivery still needs a human owning the architecture, the scope, and the merge gate, the agent is an implementation partner, not the decision-maker