Software engineer. Three-plus years shipping across any stack, product, or phase of the SDLC. Primary workflow is Claude Code + MCP — the agents build the internal tooling my team consumes as daily work product.

The five pinned repos below are clean-room extracts from a private agentic ecosystem. Each one encodes a pattern worth reusing; each ships with an integrity-check script, a claim-evidence mapping, and a green CI pipeline.

• Claude Code — skill design, hook configuration, multi-skill orchestration, sub-agent fan-out, rule-obsolescence audit discipline
• MCP — at the primitive level (Jira / Slack / Chrome DevTools / Gmail), not wrapped abstractions
• Python — for agentic tooling, pipelines, and schema-first config systems
• Next.js 16 + React 19 + TypeScript 5 strict + Tailwind v4 + MDX — for the publishing layer of the ecosystem
• SQLite + YAML schemas — the trending-store and flag-gated-config combo that makes agentic pipelines debuggable

Reference scaffold for QA-automation pipelines built on Claude Code skills + MCP patterns. 16 skill contracts, an 8-module Python scaffold demonstrating sub-agent orchestration, a 7-table SQLite trending store, and a runnable end-to-end demo over synthetic fixtures (2 boards, 30 tickets). 7 pytest tests.

Research scaffold for AI engineering with Claude Code. 15 rule files (13 persistent + 2 path-scoped); 4 pilot rules carry a **WHY:** + **Retire when:** audit-tag discipline so prompts decay cleanly as models improve. 21 skills. Concentric-loop pedagogy pinned per-node to named practitioners (Chip Huyen, Eugene Yan, Hamel Husain, Jeremy Howard, Sebastian Raschka, Andrej Karpathy, Julia Evans).

Chronicle editorial format + wiki-as-RAG graph shape for LLM-powered research labs. 2 sample chronicles, 5 cross-linked concept pages, a zero-dependency Python retrieval example. Designed so a research lab's session history becomes a retrievable corpus without any vector-DB infrastructure.

Static publisher for short research notes. Next.js 16 + React 19 + TypeScript 5 strict + Tailwind v4 + MDX + Schema.org JSON-LD. 5 static routes (3 MDX posts + home + 404). Fresh npm run build verified clean; JSON-LD confirmed in the rendered HTML output.

HANDOVER + SYNC inter-repo protocol for multi-agent AI systems. Specification + synthetic examples (3-entry HANDOVER log + 4-decision SYNC log) + stdlib-only conformance validator. The validator caught a real spec/example inconsistency during build; fix is documented in the commit history as a conformance-loop demonstration.

The private ecosystem that produced these extracts has four roles. claude-code-mcp-qa-automation is the odd one out — extracted from a separate internal-tooling track (production QA agents at work), not from the coaching ecosystem.

                   ┌──────────────────────────────────┐
                   │       Private ecosystem          │
                   │  (four roles, four repos)        │
                   └──────────────────────────────────┘
                                   │
         ┌──────────┬──────────────┼──────────────┬──────────┐
         ▼          ▼              ▼              ▼          │
     Partner    Observer       Publisher       Commons       │
   (coaching  (chronicles    (static site    (inter-repo     │
    + rules    + wiki)        rendering)      protocol)      │
    + skills)     │              │               │           │
         │        │              │               │           │
         ▼        ▼              ▼               ▼           │
 claude-code-  llm-rag-    nextjs-16-mdx-  claude-multi-     │
 agent-skills- knowledge-  research-       agent-protocol    │
 framework    graph        publisher                         │
                                                             │
                                                             ▼
                                           ┌──────────────────────────────┐
                                           │  Separate internal-tooling   │
                                           │  track (work: QA automation) │
                                           └──────────────┬───────────────┘
                                                          │
                                                          ▼
                                            claude-code-mcp-qa-automation
                                            (sanitised clean-room extract)

Each public repo is the shape — not a full copy — of what runs privately. The pattern travels; the project-specific content stays private.

Every repo enforces the same pre-push contract:

• Gate 0 — claim-evidence mapping. Every number / named capability in the README maps to a file or command that evidences it. Unverified rows fail CI.
• Gate 1 — no-hype-language audit on public docs. A fixed deny-list of adjective-driven standards (documented per-artifact in the integrity scripts); every hit must be operationalised into a measurable replacement or deleted.
• Gate 2 — clone-and-run. Fresh clone, follow the README, advertised output lands.
• Gate 3 — documentation completeness. Standard README shape, every link resolves.
• Gate 4 — identifier grep returns zero hits against a private-name exclusion list.
• Gate 5 — secret grep returns zero hits.
• Gate 6 — CI on every push runs the integrity check + hype-word audit.
• Gate 7 — fresh-eye read. Diary-shaped, aspirational, or keyword-stuffed sentences get rewritten before ship.

The scripts/integrity-check.sh in each repo runs Gates 0, 1, 4, 5 + artifact-specific verification locally before any push. CI runs the same script on every commit.

bhandari.aman0101@gmail.com