← Reports
← Back to The Den
decision

Skills & Workflow Architecture Review

Generated 2026-04-06 21:24  ·  PKA / Lucienne

The Problem

PKA has 100+ skills across global (~/.claude/skills/) and plugin (superpowers) sources. Elmar shouldn't need to remember which skill to call or in what order. The system should route automatically.

Current Architecture

LayerWhatCountStatus
Global skillsCustom skills built over months in CoWork/PKA108Active, tuned to Elmar's workflow
Superpowers pluginGeneric workflow skills from claude-plugins-official14Active, mostly overlaps with custom skills
Retired skillsArchived in _retired/ with keyword index39Lazy-loaded on demand
gstack skillsBrowse, QA, ship, review, design — from gstack plugin~40Active, some overlap

Superpowers vs Custom Skills — Full Overlap Map

Superpowers SkillCustom EquivalentWhich Is Better
brainstorming/office-hoursCustom — adds YC forcing questions
writing-plans/autoplanCustom — chains 4 review types
executing-plansDev-loop patternCustom — 4-agent team with QA
dispatching-parallel-agentsLucienne dispatch modelCustom — tier-based routing
subagent-driven-developmentDev-loop patternCustom — includes council review
systematic-debugging/investigateCustom — 4-phase with root cause rule
test-driven-development/qaCustom — browser-based, iterative fix
requesting-code-review/review + /second-opinionCustom — multi-model council
receiving-code-reviewBuilt into dev-loopCustom — reviewer agent in the loop
finishing-a-development-branch/shipCustom — includes version bump, changelog
verification-before-completionCLAUDE.md global ruleTie — same concept
using-git-worktreesAgent tool isolationTie — built into Claude Code
writing-skills/skill-evolverCustom — auto-detects patterns
using-superpowersSuperpowers wins — routing enforcer

The Routing Solution

Elmar describes what he wants in plain language. The system routes automatically:

Elmar Says...Auto-InvokesWhat Happens
"Build X" / "Add feature X"/autoplan → dev-loopFull pipeline: brainstorm → design → review → build → QA → council
"Fix this bug"/investigate4-phase debugging with root cause rule
"Is this a good idea?"/office-hoursYC-style forcing questions + design thinking
"Review this" / "Check my work"/review + /second-opinionCode review + multi-model council
"Test this"/qaBrowser-based QA, iterative bug fixing
"Ship it"/shipVersion bump, changelog, PR, push
"Research X"/deep-researchMulti-source parallel research

7 entry points cover 90% of work. The other 100+ skills are specialists called within these workflows.

Luci's Skill-Driven Tier (NEW — deployed 2026-04-06)

The same routing logic now runs on Luci for autonomous MC ticket processing:

TierWhenHow
Tier 1: One-shotSingle file, clear fixDirect fix, update ticket
Tier 2: Skill-drivenMulti-file, clear requirementsautoplan → dev-loop → QA → council → PR
EscalateDesign choice, ambiguous, stuck after 2 triesMC ticket + Telegram to Elmar

Tested end-to-end with MC-190: Luci picked up a LegalMind ticket, fixed 9 debug console.logs, pushed branch — all in 2 minutes.

Larry Enhancement (deployed 2026-04-06)

Larry's nightly LegalMind review now classifies findings by complexity:

Role of Superpowers Plugin

Keep it enabled. Superpowers' "using-superpowers" meta-skill is the one piece that doesn't have a custom equivalent — it acts as the routing enforcer, pushing Claude to check for applicable skills before every response. Custom skills take precedence per CLAUDE.md priority rules, but superpowers ensures they actually get invoked.

Vibe Kanban (parked)

Evaluated and installed locally for SafairBru + LegalMind. Decided to invest in Luci's skill-driven tier first. VK creates a laptop dependency that defeats the "work while you sleep" advantage. Revisit only if Tier 2 proves insufficient.

Decisions