~/.claude/ — harness engineer
Code is ephemeral. The harness is leverage.
I build the scaffolding around the LLM — slash-commands, agents, eval harnesses, MCP server choices, hook configuration — rather than the model itself. The footprint shows up in three places: custom Claude Code skills + agents I’ve authored, Lina Lab (a production prompt-evaluation engine), and AHKMS (a webhook-driven AI orchestration pipeline).
Custom skills authored
43 user-level skills in ~/.claude/skills/ (ground-truthed via ls), grouped below. A separate plugin-namespaced set of HeyLina rituals lives outside skills/ — see the strip below the grid.
8
Spec-driven development
/spec-* pipeline from req → design → tasks → impl → verify
/spec-req
/spec-design
/spec-tasks
/spec-impl
/spec-verify
/spec-debate
+ 2 more
11
PR review & repo hygiene
Code review, branch ops, repo cleanliness
/qa-review
/address-review
/review-and-comment
/create-clean-tree
/wrap
/merge-prune
+ 5 more
8
Engineering discipline
SOLID, TDD, architecture, shipping
/solid
/architect-senior-engineer
/improve-codebase-architecture
/tdd
/prototype
/pragmatist
+ 2 more
11
Diagnostics & dialogue
Grill, fix, think, zoom
/diagnose
/fix
/fix-failing-tests
/triage
/grill-me
/grill-with-docs
+ 5 more
5
Content & specialists
Editing, compression, specialist agents
/edit-article
/caveman
/synthesiser
/ui-ux
/heylina-pptx
HeyLina rituals + capture (plugin-namespaced)
Live as `anthropic-skills:*` rather than in `~/.claude/skills/`
14
/dos-eod/dos-reflect/dos-inbox/dos-braindump/heylina-notion/capture/quick-capture/prep/whatsapp-blast/personal-style/handoff/route/explain/boss-modeCustom agents
Multi-agent patterns for harder calls. /spec-debate orchestrates five (drafter → reviewer → defender → judge → implementer) and scores against a rubric.
architect-senior-engineer
Design / refactor with senior-engineer discipline.
pragmatist
Ship-focused scope cutter when scope creeps.
synthesiser
Weighs arguments from multiple agents and recommends a call.
voice-of-user
Represents end-user needs against technical preferences.
ui-ux
Interface review and pixel-perfect refinement.
repo-organizer / archiver
Repo cleanliness specialists; feature-folder archiving.
Production harness in the wild
Lina Lab
Production eval harness · Python / FastAPI on Railway
Versioned prompt catalog · LLM-as-judge with provenance (judge_type, judge_model, judge_prompt_version, judge_rater_id) · variant experiments · multi-scope rubrics · prompt soft-delete + version pinning · Slack-notified promotion pipeline · clinical-safety risk classification.
AHKMS
Agentic-pipeline orchestration · webhook-driven
Capture → workflow webhook fires Express worker → AI extraction → PARAMPS classification → human-in-the-loop review → derived artifacts with lineage tracking. Not "agents" in the SDK sense — but agentic orchestration in the operational sense.
Harness configuration & ops
dotclaude repo — versioned ~/.claude/ config: skills, agents, settings, hooks. Treated as a portfolio artefact, not a scratchpad.
Pre-push gates — .husky/pre-push runs typecheck + lint + test before push. Claude-Code hooks block destructive git commands (push, reset --hard, branch -D).
MCP server choices — Notion (multiple workspaces), Slack, Gmail (forked from gongrzhe), Supabase, Vercel, computer-use (native Windows control), Claude-in-Chrome (browser DOM), PostHog, Composio (auth broker), Dex (CRM). Each MCP has a job.
AI co-author footprint at scale — Co-Authored-By: Claude Opus 4.7 on most recent commits. 2,391 commits, 1.12M lines added across 44 repos.