A .claude/ framework for AI-engineering workflows with Claude Code. 15 rule files, 21 skills, concentric-loop pedagogy, rule-obsolescence audit framework.

A reusable .claude/ directory for an AI-engineering workflow with Claude Code, covering: Python fluency → production APIs → classical ML → deep learning → LLM apps → RAG → agents → production deployment.

The four things encoded here that most .claude/ setups miss:

1. Rule-obsolescence audit discipline. Every persistent rule can carry a **WHY:** tag (what default model behavior the rule corrects) and a **Retire when:** tag (the observable condition under which the rule no longer earns its presence). When a frontier model ships, compare task performance with the rule loaded vs without. If the new model handles the default correctly, the rule retires. 4 rules demonstrate the pattern today (agentic-engineering.md, concentric-loop.md, scale-position.md, session-capture.md).
2. Concentric-loop pedagogy. Every concept travels: analogy → code → OS / network → hardware / math → analogy (return). The return is mandatory. See .claude/rules/concentric-loop.md. Pinned per-node to working practitioners (Chip Huyen, Eugene Yan, Hamel Husain, Jeremy Howard, Sebastian Raschka, Andrej Karpathy, Julia Evans).
3. Session-capture dual-artifact contract. Every teaching exchange produces a chronological session entry AND an atomic wiki node with cross-links. Session log feeds narrative; wiki feeds retrieval. See .claude/rules/session-capture.md.
4. Scale-position axis (0-10). A single hirable-mastery dimension anchored to named practitioners at each point (Simon Willison 2 → Eugene Yan 3 → Chip Huyen 5 → Karpathy 6 → Raschka 7). Sets what "good enough" means so depth-vs-speed trade-offs resolve consistently. See .claude/rules/scale-position.md.

• Claude Code .claude/rules/ for persistent behavioral contract
• Claude Code .claude/skills/ (v2 skill format with frontmatter) for session-mode behaviors
• Bash integrity-check script + GitHub Actions CI for pre-push gate enforcement

No runtime dependencies beyond Claude Code itself.

claude-code-agent-skills-framework/
├── CLAUDE.md                  # .claude/ entry point
├── ROADMAP.md                 # 5-project curriculum map
├── .claude/
│   ├── rules/                 # 15 rule files (13 persistent + 2 path-scoped)
│   │   ├── agentic-engineering.md      ★ WHY-tagged
│   │   ├── code-standards.md
│   │   ├── concentric-loop.md          ★ WHY-tagged
│   │   ├── context-management.md
│   │   ├── exercises.md                (path-scoped: topics/**)
│   │   ├── explorer-builder.md
│   │   ├── knowledge.md                (path-scoped: knowledge/**)
│   │   ├── math-foundation.md
│   │   ├── partner-identity.md
│   │   ├── problem-first.md
│   │   ├── progressive-assessment.md
│   │   ├── scale-position.md           ★ WHY-tagged
│   │   ├── session-capture.md          ★ WHY-tagged
│   │   ├── session-design.md
│   │   └── systems-thinking.md
│   └── skills/                # 21 skills, each with SKILL.md + frontmatter
│       ├── adr-capture/             # Architectural Decision Record capture
│       ├── adversarial-review/      # Chief-architect-style stress test
│       ├── architect-first/         # Forces planning before code
│       ├── blog-draft/              # Long-form technical writing helper
│       ├── brainstorm/              # Ideation + wiki-knowledge-builder
│       ├── code-review/             # Coach-style code review
│       ├── daily-checkin/           # Session-start context sync
│       ├── debug-systems/           # Three-layer debugging protocol
│       ├── evolution-log/           # Topic-completion growth log
│       ├── faang-loop/              # Interview-prep practice loop
│       ├── learn-in-public/         # Short-form + long-form posts
│       ├── mock-assessment/         # Whiteboard-test dry run
│       ├── os-lens/                 # dis / strace / /proc inspection
│       ├── partner/                 # Core coaching-identity behaviors
│       ├── progress-update/         # PROGRESS.md session entry
│       ├── review-systems/          # Systems-aware code review extension
│       ├── spaced-review/           # FSRS-style concept review deck
│       ├── system-check/            # Weekly system-health audit
│       ├── systems-trace/           # Three-layer concept introduction
│       ├── trend-scout/             # AI-engineering trend digest
│       └── weekly-reflection/       # Friday written reflection
├── docs/
│   └── claim-evidence.md      # Gate 0: every README claim evidenced
├── scripts/
│   └── integrity-check.sh     # Gates 0, 4, 5 + artifact-specific checks
└── .github/workflows/ci.yml   # Runs integrity check + hype-word audit

# Clone
git clone https://github.com/aman-bhandari/claude-code-agent-skills-framework.git
cd claude-code-agent-skills-framework

# Drop .claude/ into your own project (or use this dir as the root)
# Claude Code reads .claude/rules/*.md on every session and loads skills on demand

No npm install, no pip install. This is configuration only.

1. Start Claude Code from this directory: claude (or claude-code).
2. Claude Code reads CLAUDE.md + .claude/rules/*.md automatically on session start.
3. Invoke a skill by name from Claude Code's slash menu — e.g., /daily-checkin, /brainstorm, /architect-first, /review-systems.
4. Adapt the role vocabulary in CLAUDE.md (Coach/Student/Observer/Publisher) to your project.
5. Customize rules for your own curriculum. Keep the WHY-tag + Retire-when discipline so rules decay cleanly as models improve.

bash scripts/integrity-check.sh

Runs:

• Gate 0 — every claim in docs/claim-evidence.md marked verified (✅)
• Gate 4 — zero private identifiers (project-internal names, client domains, real individuals)
• Gate 5 — no secrets (API keys, tokens, passwords) outside .env.example
• Artifact-specific — rule count = 15, skill count = 21, every skill has SKILL.md, ≥4 rules carry WHY + Retire audit metadata

CI runs the same gates on every push via .github/workflows/ci.yml.

• claude-code-mcp-qa-automation — end-to-end QA automation built on Claude Code + MCP patterns
• llm-rag-knowledge-graph — chronicle editorial format + wiki-as-RAG graph shape
• nextjs-16-mdx-research-publisher — static publisher for research labs
• claude-multi-agent-protocol — HANDOVER + SYNC inter-repo protocol

MIT © 2026 Aman Bhandari. See LICENSE.