Skills Inventory
skills inventory Link to heading
Comprehensive inventory of agentic-development and Claude Code competence, anchored on ten current production works.
Anchor works Link to heading
| # | Work | What it is |
|---|---|---|
| 1 | JOBBER (v3 → v4) | Go monorepo automating job-listing collection, ingest, and curator-console web UI; msgvault MCP server wired in; Vault-sourced Postgres creds; ~2,937 LOC Go in v4 |
| 2 | md-util (mdu) | SHACL-first data-governance CLI (Go, ~10,879 LOC, 13 packages, 8 subcommands, v0.8.0); never moves rows, never executes live DDL |
| 3 | Nate B. Jones 5-Level Prompting | Passed the five-level test — external validation of prompt competence |
| 4 | Phone UI | Mobile front-end built with Claude Code — cross-modality fluency |
| 5 | Southwall + 67 codified skills | Homelab IT maintenance with Claude + Oz — kafka-topic-admin, vault-secret-rotate, pggold-schema-manage, homelab-health-check, netbox-inventory-sync, southwall-dotfiles-sync, and more |
| 6 | Claude Cowork + Linear | AI-assisted CI/CD and PM — Linear MCP for ticket lifecycle, Cowork for collaborative agentic delivery |
| 7 | Temporal workflows | southwall-streams, metro-streams, training-streams — deterministic workflow code authored agentically |
| 8 | Det vs non-det trade-off framework | Architectural distinction between where determinism is mandatory (workflows, DDL, money paths) and where non-determinism is acceptable (synthesis, design, exploration) |
| 9 | Organized prompt library | 8 promptsets, 377-line XML prompt schema, 173+ lines of session logs |
| 10 | LMTRAIN | First locally trained small model (nanoGPT, Obsidian-vault character-level checkpoint on MPS) |
Claude Code as a platform Link to heading
- Custom skills. 67 versioned, scoped capabilities across ops, design, QA, debug, deployment
- Scoped permission profiles. Per-project
.claude/settings.local.jsonallowlisting specific binaries (go test,go build,mdu *, curated git commands) - Persistent memory layer. Slug-indexed
.mdmemory files plusMEMORY.mdindex; reference memories kept hot across sessions - MCP servers in production use. Linear (92-ticket cycle), msgvault (JOBBER mail ingest), MotherDuck, claude.ai-hosted (Gmail, Drive, Calendar). Custom MCP wiring, not vendor-default
- Status-report discipline.
.context/status_YYYY-MM-DD_HHMM.mdafter each step — auditable agent provenance - Hooks for session lifecycle. SessionStart hooks for behavior customization (e.g., Ponytail mode)
Skill ladder Link to heading
Must Have — carried Link to heading
| Skill | Evidence |
|---|---|
| Claude Code at platform level | 67 skills, scoped permissions, persistent memory, MCP, hooks |
| Production agentic builds | md-util (10,879 LOC), JOBBER (~2,937 LOC), phone UI, Temporal streams |
| MCP integration | Linear, msgvault, MotherDuck, claude.ai MCPs |
| Prompt architecture (multi-pattern, meta-prompting) | Nate B. Jones 5-level pass; 8 promptsets; 377-line XML schema |
| HDD–SDD–BDD methodology | practice-docs, interactive Go CLIs, 20,350-line book outline |
| Det vs non-det trade-off command | md-util governance-only; Temporal for deterministic side |
| Schema-first governance for AI-touched data | SHACL 1.2 Turtle, mdu: vocab, multi-target emit |
| Ops as agentic skills | Southwall + Oz + 67 skills |
| Frontier-LLM inference orchestration | Claude Code production loops; Cowork + Linear |
Should Have — carried Link to heading
| Skill | Status |
|---|---|
| Vector / memory R&D (Qdrant) | Carried |
| Workflow orchestration (Temporal) | Carried |
| Curriculum / enablement design | Carried (Hyperion Master Curriculum + AI 5-section program) |
| Local-first homelab AI environment | Carried (Vault, Consul, Kafka native on Gold, direnv) |
| Local model training | Carried (LMTRAIN / nanoGPT) |
| Cross-modality delivery (CLI, web, mobile) | Carried |
| Lineage / observability for AI artifacts | Carried (md-util DuckDB lineage) |
Nice to Have — gaps Link to heading
| Skill | Status |
|---|---|
| Published OSS artifact for the AI work | Gap (see roadmap) |
| Conference talk on agentic methodology | Gap |
| Anthropic builder partner credential | Gap |
| Formal eval rig (Inspect / Promptfoo / homegrown) | Gap |
| RLHF / fine-tuning at scale | Gap (deliberate; not pursuing) |
| Multi-agent fan-out (LangGraph / AutoGen / Crew) | Gap (Claude Code pipelining covers most use cases) |
Fit review Link to heading
Strong signals
- Volume and variety of production agentic work — ten anchor works spanning CLI, web, mobile, workflow, ops, PM, and model training
- Operating Claude Code at platform scale — 67 codified skills is a rare signal; custom MCP wiring (JOBBER msgvault) rarer still
- Externally verified prompt skill — Nate B. Jones 5-level pass is a third-party benchmark
- Methodology, not just usage — HDD–SDD–BDD codified as a named framework
- Hands-on with the model end — LMTRAIN closes the “only an inference user” objection
- Strategic command of det vs non-det — the senior-engineer judgment call
- 30+ years of DW/BI depth underneath — reviewers don’t have to teach fundamentals
Soft signals (gaps)
- No public OSS for the AI work — all anchor works are private
- No frontier-lab employment line — Hyperion / Vertica / Full360 anchor the resume; Anthropic / OpenAI / Databricks do not
- No conference talk on agentic methodology — presentation-ready but unpresented
- No formal eval rig named (mitigation: build one against md-util’s dataset-batch loop)
Tools Link to heading
Claude Code (skills, custom MCP, hooks, scoped permissions, persistent memory) · Claude Cowork · Anthropic API · MCP (Linear, msgvault, MotherDuck, custom) · Qdrant · Temporal.io · SHACL 1.2 · DuckDB / DuckLake · Apache Iceberg + AWS Glue · Go (Cobra, pgx/v5, aws-sdk-go-v2, go-duckdb) · Python 3.11 · nanoGPT / PyTorch / MPS · Vault · direnv · Linear · GitLab · OpenMetadata.