🐝 Hermes Kanban Swarm#

Ultra-condensed quick-start for turning broad Funday goals into a persistent, delegated, evidence-gated agent swarm.

Golden rule: Hermes orchestrates. Kanban remembers. Skills define the method. Workers execute. Verification blocks lies.

⚡ When To Swarm#

Use a Kanban swarm when work is:

Trigger	Why
🎮 “Make this game perfect / more fun”	Needs audits, fixes, proof
🐛 “Still broken”	Needs root-cause lanes + verification
🧠 Multi-skill work	FunOps + Svelte + deploy + UI
⏳ Long-running	Must survive restart / Discord thread drift
🧪 Needs proof	Separate verifier card prevents fake done
🧑‍🤝‍🧑 Needs specialists	Parallel profiles finish faster

Do not swarm tiny one-shot questions. Answer directly.

🧬 Stack Map#

User intent
  → Hermes routes
  → Skill chooses method
  → Kanban board stores graph
  → Worker profiles execute cards
  → Verifier card proves completion
  → Memory/docs preserve learning

Layer	Role
🪽 Hermes	Orchestrator, not grinder
📋 Kanban	Persistent dependency graph
🧩 Skills	Workflow DNA (`fun-audit`, `funday-play-shell`, `svelte-code-writer`)
🤖 Profiles	Focused workers
✅ Verification	Blocks false “done”
🧠 Memory/docs	Makes the next loop smarter

🧭 First 60 Seconds#

# 1. Know who can work
hermes profile list

# 2. Create an isolated board
hermes kanban boards create orbit-snatch-fun-pass --name "Orbit Snatch Fun Pass"
hermes kanban boards switch orbit-snatch-fun-pass

# 3. Verify swarm skills are enabled
hermes skills list | rg 'hermes-kanban-swarm|fun-|funday-play-shell'

Never invent assignees. Unknown profile names silently strand cards.

🎮 Default Game Perfection Graph#

Use this for:

"make <game> more fun"
"make <game> perfect"
"run a delegated game improvement swarm"

flowchart LR
  A[deploy-health<br/>funday-play-shell] --> E[synthesize findings]
  B[fun-audit<br/>Fun Stack + moment map] --> E
  C[ux-quality<br/>audit / critique] --> E
  D[browser-playtest<br/>3 personas] --> E
  E --> F[fun-plan<br/>top 3 fixes]
  F --> G[fix blocker 1]
  F --> H[fix blocker 2]
  F --> I[fix blocker 3]
  G --> J[fun-verify]
  H --> J
  I --> J
  J --> K[polish / ship summary]

Hard gate: If deploy-health finds blank viewport, JS 404, stale build, or broken Nakama handler, fix platform first. Do not score fun on a broken launch.

🧩 Skill Routing#

Work	Skill
Swarm graph / delegated board	`hermes-kanban-swarm`
Game feel diagnosis	`fun-audit`
Feel improvement plan	`fun-plan`
Prove feel improved	`fun-verify`
Blank route / dock / drawer / deploy	`funday-play-shell`
UI quality / a11y / perf	`audit`, `critique`, `optimize`
Svelte implementation	`svelte-code-writer`, `frontend-development`
Juice / animation	`delight`, `animate` after diagnosis
Final detail pass	`polish` after verification

📋 Perfect Card Shape#

Cards must be scoped and readable in 30 seconds.

## What to build
Run evidence-based FunOps audit for orbit-snatch.

## Steps / Key files
1. Use `fun-audit`.
2. Play 3 personas: novice, competent, greedy.
3. Save report to `docs/fun-ops/audits/YYYY-MM-DD-orbit-snatch-fun-audit.md`.

## Acceptance criteria
- [ ] Scores cite evidence or are marked `?`
- [ ] Top 3 findings ranked by `(Impact × Confidence) / Effort`
- [ ] No code changed

## Blockers
- If game is blank, route to `funday-play-shell` first

🛠 CLI Pattern#

# Parent cards: independent, run in parallel
hermes kanban create "audit: orbit-snatch deploy health" \
  --skill "funday-play-shell" \
  --assignee "<existing-profile>" \
  --body "<condensed card body>" \
  --json

hermes kanban create "audit: orbit-snatch game feel" \
  --skill "fun-audit" \
  --assignee "<existing-profile>" \
  --body "<condensed card body>" \
  --json

# Child cards: born gated with --parent
hermes kanban create "plan: orbit-snatch top 3 feel fixes" \
  --skill "fun-plan" \
  --assignee "<existing-profile>" \
  --parent "<fun-audit-task-id>" \
  --json

CLI gotcha: task title is positional:

hermes kanban create "Task title"   # ✅
hermes kanban create --title "Task" # ❌

🔒 Non-Negotiable Gates#

Gate	Rule
👤 Profile gate	Discover real profiles first
🧠 Plan gate	No implementation before plan/approval
🔗 Dependency gate	Use `--parent`; prose does not gate work
🧪 Evidence gate	No audit scores without evidence
✅ Verify gate	No “done” without verifier card
🎮 Fun gate	No “more fun” without `fun-verify`
🚢 Deploy gate	Build changes require restart/smoke
🧹 Scope gate	Top 3 fixes; backlog the rest

🔁 Bounded Perfection Loop#

Loop, but do not spiral.

Sense → Decide → Act → Verify → Remember → Repeat only if valuable

Continue if:

🔴 blocker remains
verification failed
new evidence reveals high-impact issue
user asks for another pass

Stop if:

verify passes
remaining work is low-impact polish
risk exceeds reward
scope is complete

Each loop must produce one artifact: audit, plan, fix, verify record, or blocker.

🧠 Mental Model#

Bad swarm:

One giant "perfect the game" card
→ one worker edits everything
→ no evidence
→ fake done

Good swarm:

parallel audits
→ synthesis
→ top 3 plan
→ scoped fixes
→ independent verification
→ memory/docs update

🏆 Ultimate Compression#

Swarm perfection = correct decomposition + hard gates + persistent state + proof before claims.

For games:

deploy-health + fun-audit + ux-audit + playtest
  → synthesis
  → fun-plan
  → top-3 delegated fixes
  → fun-verify
  → polish / ship / loop

Use many agents only when the graph is clean. The power is not the swarm size. The power is the dependency-safe board.