## 12. THE SKILL SYSTEM

Skills are reusable instruction files the owner attaches to sessions and worker briefs.
The skill-usage miner scanned **1,194 transcripts** for invocations and skill-bearing
briefs. The distinctive finding: the most-used skills are not generic tools — the
presentation-worthy tier is **self-authored** (`grill-me`, `loop-eng-format`,
`mem0-memory`, `caveman`, and the principle skills), each codifying
one of the owner's own rituals so it never has to be re-explained.

### 12.1 Ranked skill inventory (by distinct sessions)

Self-authored / self-codified skills marked ★.

| # | Skill | Origin | Sessions | What it does | When invoked | Verbatim invocation |
|---|-------|--------|----------|--------------|--------------|---------------------|
| 1 | `review-bugbot` | Cursor built-in | **129** | Code-reviewer subagent: correctness, regressions, logic bugs against the base-branch diff | Pre-merge gate | "Before merge-prep: /review-bugbot (whole branch) + /thermos, fix all findings, second scoped /review-bugbot on the fixes" |
| 2 | `thermos` ★ | Custom plugin | **55** | Two parallel auditors (bugs/security/devex + structure quality, incl. the 1,000-line max-file-size rule and spaghetti detection) synthesized into one deduplicated verdict | Pre-merge gate | "/review-bugbot (whole branch) then /thermos (bugs+quality audits)" |
| 3 | `browse` | gstack suite | **41** | Fast headless Chromium for UI QA, live-route dogfooding, screenshot proof (~100ms/action) | Mid-build / prove station | "With …/browse on 'http://localhost:3000/dev-preview?view=signoff': capture proof" |
| 4 | `heroui-react-pro` | Custom guide | **38** | Enforces HeroUI v3 conventions: compound dot-notation anatomy, onPress-not-onClick, MCP component lookup | Mid-build (FE briefs) | "…/heroui-react-pro/SKILL.md — HeroUI v3 usage: compound dot-notation anatomy, onPress not onClick, semantic variants" |
| 5 | `heroui-pro-design-taste` | Custom design profile | **37** | 78 design-taste principles across 10 categories (spacing, typography, caution-yellow budgets, card anatomy) distilled from the owner's own feedback | Mid-build (FE briefs) | "78 design-taste principles; governs composition/quality calls alongside DESIGN.md" |
| 6 | `principle-redesign-from-first-principles` ★ | Custom principle | **33** | Forbids bolt-on parallel layers; forces rethinking the subsystem as if the requirement existed on day one | Kickoff / architecture | "/principle-redesign-from-first-principles Consider the GREENLIGHT Steps integration… from the first principle view" |
| 7 | `caveman` ★ | Custom prompt register | **27** | Strips filler and pleasantries to cut token burn; code, diffs, and verdicts stay byte-exact | Standing worker constraint | "report in caveman mode (full). Compress prose only — file excerpts and diffs stay EXACT" |
| 8 | `mem0-memory` ★ | Custom self-authored | **26** | Session-start memory retrieval + hook-gated 24h curation checkpoints distilling durable lessons | Startup / 24h curation | "Run the mem0-memory skill as a curated checkpoint covering the work since the last checkpoint" |
| 9 | `grill-me` ★ | Custom self-authored | **19** | Socratic elicitation: agent interviews the human in front-loaded multiple-choice batches, recommended option first | Kickoff / scoping | "let's discuss in details, use grill me to get all the answers" |
| 10 | `loop-eng-format` ★ | Custom self-authored | **18** | Compiles a slice plan (`SETUP-X.md`) into a fresh-chat execution doc (`START-X.md`) with maker/checker split, circuit breaker, state file | Kickoff / slice start | "/loop-eng-format tasks/nomos-v1/SETUP-V2.8.md" |
| 11 | `never-block-on-the-human` ★ | Custom principle | **16** | Authorizes autonomous execution of reversible decisions and verified fixes; human pauses only at irreversible gates | AFK modifier | "Owner is AFK for the rest of this run. Adopt never-block-on-the-human posture for everything already in your brief." |
| 12 | `02-vulnerability-scanner` | Custom security | **15** | Static security audit: secrets leakage, unsafe queries, injection, auth bypass | Pre-merge gate | "…runs /review-bugbot (whole branch) then /thermos then vulnerability scan" |
| 13 | `deslop` | Plugin (cursor-team-kit) | **14** | Strips AI code slop from branch diffs: narrating comments, reflexive try/catch, `any` casts, nested wrappers | Pre-PR hygiene | "/deslop Deslop the code base in this branch." |
| 14 | `finishing-a-development-branch` | Plugin (Superpowers) | **14** | Structured branch closeout: verify suites → PR/merge options → curated squash message → residue cleanup | Pre-merge / integration | "/finishing-a-development-branch" |
| 15 | `interrogate` ★ | Custom review | **8** | Spawns multiple frontier models with identical rubrics to adversarially challenge a design's assumptions | Kickoff / deep review | "plan via /principle-redesign-from-first-principles + /interrogate before build" |

### 12.2 The four evidenced skill chains

1. **Pre-merge multi-lens audit.** — Phase 1 (parallel): `/review-bugbot` (whole branch)
   + `/thermos` (bugs/security + structure/quality) + `/02-vulnerability-scanner`, run as
   read-only auditors with commits frozen. Phase 2: synthesize one severity-sorted table →
   attach `never-block-on-the-human` → fix all verified findings (blocking → simple →
   complex) without pausing. Phase 3: scoped `/review-bugbot` on the fix commits only →
   `/deslop` → `/finishing-a-development-branch` for the curated squash message.

2. **Slice kickoff and specification.** — `grill-me` (multiple-choice elicitation) →
   `/principle-redesign-from-first-principles` + `/interrogate` (multi-model adversarial
   stress-test) → draft `SETUP-V2.x.md` (vertical tracers, DONE-WHEN predicates — explicit
   exit criteria per deliverable) →
   `/loop-eng-format` compiles it into `START-V2.x.md` with an external `LOOP-STATE-*.md`
   tracker, maker/checker split, and circuit breakers.

3. **Frontend builder steering.** — Worker briefs mandate `heroui-react-pro` (verified
   component API via MCP, no hand-rolled primitives) + `heroui-pro-design-taste` (78
   principles) + `caveman` (token-efficient reporting); an *independent* checker then
   drives `browse` against the live app for screenshot proof —
   the builder never grades its own integration.

4. **Cross-session memory curation.** — A global hook runs `mem0-memory` at session start
   and injects `MEM0_CONTEXT_READY:<hash>`; a background curation pass (at most once per
   24h, hook-enforced) distills transcript deltas into mem0 + `AGENTS.md` autonomously.