Multi-agent system · agentskills.io

An agent that builds — and verifies — its own game. PawnSim colony-sim · two media skills · human-in-the-loop ops · $0 runtime · EN + KO

AI-Powered Self-Evolving Autonomous

The active work is PawnSim — a colony-sim vertical slice the agent builds and play-tests in a tight loop, where the headline isn't features but verification: code existing ≠ verified working. A production media skill (music-video) ships alongside a parked one (job-hunt), every mechanical stage stays local at $0 runtime cost, and the repo audits itself on every commit.

Every commit is gated by a 22-scenario input-level repro test — synthesized clicks through the player's own UI path, asserting the effect, not just that the click landed.
Long unattended soaks are judged by an isolated grader sub-agent against a written rubric — it sees only evidence (screenshots + raw logs), never the author's intent.
Two media skills: music-video (a track → a 9:16 short with phrase-aware cuts + 23 vintage shaders) and job-hunt (one seed keyword → a deduplicated Korean job-board digest — parked, KR boards now block scraping).
Mechanical stages stay local (Unity batchmode, aubio, whisper.cpp, ffmpeg); creative stages opt into Claude under the operator's existing subscription.

See PawnSim How it works View on GitHub See the output Try it in 60 seconds

Unattended 16 in-game-day colony soak — agent-built and agent-verified. The loop shown (stockpile → housing → farming → logging → mining) is gated by the 22-scenario input-level repro test and confirmed by isolated-grader rubric verdicts, not the author's claims.

100+ mission outputs shipped across 6 mission types
1+1+4 production skill · meta-skill · prototypes music-video · job-hunt (parked) · game-dev-agent
23 ffmpeg shaders 3 stages · genre-routed
0 runtime API tokens local: ffmpeg · whisper · ollama
22 scenario commit gate input-level repro + assertions
27 subagent definitions 6 core + 13 game + 5 content + 3 judges
3 audit trigger layers commit · anomaly · schedule
MIT license · EN + KO dual-track v0.4.0 · public

PawnSim · build & verify

PawnSim — built and verified by the agent

The most active surface right now is PawnSim, a top-down colony-sim vertical slice (Unity 6000.0.75f1 LTS). Every sprite (a full 32px art generation — 3-direction walk/work pawn sheets, animals, terrain, furniture, all procedurally generated), every scene, and every C# system is CLI-scaffolded by the game-dev-agent meta-skill with no manual Unity Editor work — the whole .exe is reproducible from the command line. Colonists chop / mine / farm / cook / haul / build / research / fight under a utility-AI; an AI Director schedules raids on a jittered clock; the player drafts pawns and paints build + designation orders.

Two verification gates diagram. Gate 1 — Input-level repro gate, fires on every commit, 22 scenarios: synthesized clicks drive the same UI path a player uses, with effect assertions ('the click placed a designation', not just 'the click landed'). Gate 2 — Isolated grader sub-agent, judges long soaks against a written rubric and sees only screenshots and raw logs, never the author's intent. Below: three bugs the grader caught that self-review missed (a silent harness blind spot that voided every designation, a food-rich colony starving to death, a permanent mental-break freeze), and the basic colony loop now machine-verified end-to-end: stockpile → housing → farm plots → logging → mining → deconstruct. — The two verification gates — README · PawnSim verification

The headline is verification, not features

The north star is "code existing ≠ verified working." Two layers carry it:

22-scenario input-level repro gate — on every commit, real synthesized clicks run through the same UI path a player uses, with effect assertions: "the click placed the designation," "the wall got built," "only the selected tree was chopped" — not just "the click landed."
Isolated-grader soak loop — long unattended soaks are judged by a separate sub-agent against a written rubric. The grader sees only evidence (screenshots + raw logs), never the author's intent, which blocks author self-grading bias.

A failing gate means no advance: fix-in-place or roll back. The basic colony loop (stockpile → housing with real indoor effects → permanent farm plots → logging → mining → deconstruct) is machine-verified end-to-end, with the grader verdicts committed alongside the fixes.

What the grader caught that self-review missed

The verdicts repeatedly surfaced defects the author's own review passed over — the honest proof the loop is worth its cost:

A silent harness blind spot that had voided every designation in earlier soaks (the verifier itself was broken — "the verifier must be verified too").
A "food-rich colony starving to death" mood-gate trap.
A permanent-mental-break colony freeze.

Honestly tracked open gaps remain — some save/load entity sub-state isn't serialized yet, and the mood economy still runs a slight deficit (the one acknowledged open gameplay problem). Full honest verification status in skills/game-prototype/README.md.

Human-in-the-loop ops

The agent doesn't run open-loop. The operator plays the build, files in-game feedback, and that feedback becomes the next batch of gated fixes — an operator → agent → verify → operator cycle. Logic changes (pawn behaviour, balance) are explicitly operator-gated: the agent drafts the spec and waits for an OK before touching them.

This is the same human-in-the-loop pattern the repo applies to its media missions: the agent does all the mechanical work, the human owns taste, money, and logic approval.

PawnSim 2026-06-12 — a walled room with door and roof shading (top-down block walls from the 32px art generation), farm plots with growth-stage crops, a stockpile filling with logs and produce, named colonists, and live resource counters. — **Basic loop verified end-to-end**
walls + roof · farm plots · stockpile · named colonists

PawnSim night — colonists asleep in three visually distinct bed tiers (sleeping spot / wood bed / fine bed) with persistent zZ markers, under the night tint. — **Night cycle**
three bed tiers · sleep state · night-tint lighting

The meta-skill that scaffolds it: skills/game-dev-agent/. PawnSim is Unity / Windows-primary (the build chain runs the Editor in batchmode); the rest of the repo is macOS / Linux.

Audience · who this is for

Who this is for

You want to see an agent verify the game it builds — not just build it. PawnSim ships with a 22-scenario input-level repro gate on every commit and an isolated-grader rubric loop for long soaks; the graders' verdicts (not the author's claims) are the acceptance record, committed alongside the fixes.
You want short-form vertical video output without writing pipeline code. Give the wizard a music file, get back a 9:16 short with beat-aligned cuts and vintage shaders. No Premiere, no After Effects, no GUI.
You want to study a working multi-agent system that doesn't pretend to be magic. Every commit is one observable step in how the system evolves; docs/audit/ records every drift the auditor catches; the quality + autonomy charts below chart whether those claims hold up over time.
You want to see a Korean job-board digest skill that respects how you actually search — currently parked. Pass --seed "Problem Solver"; the skill expands to the 26 equivalent titles companies use (FDE / Applied AI Engineer / Generalist / Founding Engineer / …) across 11 sources. The KR boards now block scraping, so it runs on mock/dry-run data only; kept in the tree rather than pretending it still ships live.
You want an agentskills.io-compliant Skill you can drop into other runtimes. Both skills work in Claude Code, Cursor, Goose, Gemini CLI, OpenAI Codex, GitHub Copilot, and ~38 other listed compatible runtimes.

If you want a SaaS that hides the pipeline, this isn't it. If you want every step as inspectable bash + open-source local tools (ffmpeg / whisper.cpp / ollama / aubio), it is.

Architecture · how it works

How it works

The scaffold is general-purpose — it doesn't force every skill through one shape. Short-form video was the v1 domain (the deliverable is visually verifiable and failure modes are quick to catch); the current development focus is the game track (the PawnSim build-and-verify loop above), driven by a game-dev-agent meta-skill. One production media skill ships today, the music-video mission; the standalone job-hunt skill is parked (KR job boards now block scraping). Earlier missions (faceless-short narration, v1 highlight / shorts-batch) remain in the tree as alternate paths.

MelonS-Agents system-at-a-glance poster: an agent that builds, plays, and verifies its own colony-sim game, portable across Claude Code, Cursor, Goose, Gemini CLI, Codex and Copilot. Stat cards: 100+ mission outputs across 6 types, 1 production skill + 1 meta-skill + 4 prototypes, 23 ffmpeg shaders, 0 runtime API tokens, 22-scenario commit gate, 27 subagent definitions (6 core + 13 game + 5 content + 3 judges), 3 audit trigger layers, MIT license / EN + KO dual-track. — The system at a glance — README · Overview

The 3-shape skill model: one skill invocation fans into three pipeline shapes, chosen by where the real work lives. Shape A (Missions-routed, primary) — a 5-agent orchestrator-driven pipeline (Orchestrator/Planner/Resourcer on Opus, Editor/QA on Sonnet) with plan.md / MANIFEST.md / qa-report.md file handoffs; example skills/music-video. Shape B (Standalone) — the skill is the implementation: HTTP → parse → format → render, skipping orchestrator and plan files; example skills/job-hunt is all curl + jq. Shape ? (open) — future skills decide per-skill via SKILL.md metadata.pipeline-source. — One invocation, three pipeline shapes — README · Architecture

— Game track · the current focus

game-dev-agent (meta-skill, Skill #3)

A Unity-focused agent that CLI-scaffolds a whole game with no manual Editor work: sprite generation (32px procedural art), C# system scaffolding, scene + prefab generation via Unity batchmode, balance tuning, audio generation, and an in-game AI Director. It drives four prototype skills that double as its empirical validation surface — PawnSim (colony-sim flagship), plus a 2D physics-merge puzzle, a wave-survival action game, and a sliding-tile number puzzle — each built faster than the last, testing the "compounding speedup" hypothesis.

The game track runs a separate roster of 13 game-domain sub-agents (director / designer / programmer / build-engineer / QA / artist / sound / narrative / specialists) on top of the 6 core media-pipeline agents, 5 for the content-shorts pipeline, and 3 judges that score a plan, a still, and a rendered cut — 27 agent definitions total in .claude/agents/.

— Media pipeline · music-video mission

Media pipeline: six subagents in a committed-file relay (3 on Opus, 3 on Sonnet). The Orchestrator (Opus) decomposes a mission and delegates down a Planner → Resourcer → Editor → QA line; each agent reads upstream artifacts (plan.md → MANIFEST.md → CHANGELOG.md → qa-report.md), never a shared chat. Resourcer runs ffmpeg / yt-dlp / whisper. A separate out-of-band Auditor (Sonnet) runs a daily 03:00 repo-wide drift / contract / cost / security audit, writing docs/audit/<date>-<focus>.md and CURRENT-ALERT.md. — Six subagents, file-based handoff — README · Architecture

— Execution graph · shorts production (LangGraph)

Shorts production graph: a measured cost bar showing video rendering takes 179 of the 189 minutes in a 26-cut short, with the gate marked right after the stills, then the stage flow plan to still round (9.0s each) to Gate 1 to human approval (interrupt) to clip round (412.3s each) to Gate 2 to assemble and legal to release package, plus the four backward edges and the blocked rail. — Shorts production graph — a gate in front of the expensive stage (412 s per cut) · rendered from the running graph

— Execution graph · game production (LangGraph)

Game production graph: the same fan-out joins on a mutex instead of a gate. Publish task to review to three parallel lanes (code, art, sound) to a Unity critical section that pins artifact paths into state, then verification on the real build, TA art review, and merge, plus the false-verification and retry rails. — Game production graph — parallel lanes join on a mutex, not a gate · rendered from the running graph

Pipeline (music-video mission)

Beat extraction. aubiotrack finds real beats; sub-beat noise rejected. Cuts land every Nth beat (default 12 — about one cut per 7.5 s at 95 BPM).
Phrase alignment. aubioonset detects drum hits. Variable per-clip setpts by mood: slow scenes 0.55×, ambient 0.70×, active 0.80×, natural 1.00× — the music drives the visual pace.
B-roll. Mood-keyword Pexels Videos API fetch; per-window selection. Demo mode bundles CC-BY Blender open-movie clips for zero-key first-touch.
Glitch micro-edits. 0.2 s reverse + 0.2 s forward jump-cut on detected drum onsets, but only on clips classified as static-camera so the frame doesn't shake during the glitch.
Vintage lo-fi shaders. Film grain + vignette + Gaussian zoom-pulse + phrase-aware pond ripple + halation bloom. All pure ffmpeg filter graphs — no GLSL, no external renderer.
Render + QA. ffmpeg 9:16 screen-fill, mission-level retry on failure.

Quality bar — 5 contracts the system now enforces

The 2026-05-22 music-video QA pass surfaced six taste directives that the prior pipeline produced quietly broken output against. Five landed as enforced contracts; the sixth is open as a research direction. Case study #9 writes the framing: the bug wasn't the renders, it was that the contracts weren't expressible in code.

A.1 — B-roll dedup registry. records/youtube/broll-used.txt (271 ids seeded). Both Pexels callers consult + append.
A.2 — Lyric vocal-onset alignment. scripts/music-video-lyric-align.sh derives LRC from plain text + audio via whisper (word-level KR, segment-level EN).
A.3 — Language anchor + QA gate. lang_anchor on every preset; scripts/music-video-qa-anchor.sh scores B-roll keywords — exit 0 PASS / 1 WARN / 2 FAIL.
B.1 — Shader vocabulary. Expanded from 4 to 23 effects across three stages, genre-aware preset routing.
C.1 — Shader restraint gating. shader_active_ratio per preset; four gate modes via MUSIC_VIDEO_SHADER_GATE.

Full changelog: skills/music-video/CHANGELOG.md.

23 ffmpeg shaders — at a glance

All pure ffmpeg filter graphs. No GLSL, no external renderer. Catalog in scripts/music-video-shaders.sh; per-genre routing in skills/music-video/data/genre-presets.yaml.

Stage 1 — first pass (2026-05-17). pond, halation, breathing, combo.
Stage 2 — genre-aware expansion (2026-05-21). scanline, chromatic_split, neon_edge, vhs, saturation_pulse, kaleidoscope, beat_burst, strobe, shake, color_burst, light_rays.
Stage 3 — quality-bar batch (2026-05-22). light_leak, duotone, vignette_pulse, paper_grain, dust_speck, posterize, trail_echo, soft_bloom.

Deliberately deferred — cel-shading / cartoon (needs GLSL / EbSynth / AI stylization). See case study #5.

— Audit & cost · the self-watching repo

Three auditor trigger layers, all out-of-band and read-only, converging on one output channel. L1 — post-commit hook fires on drift-risk paths only, runs audit-run.sh contract. L2 — mission-anomaly poll every 15 minutes runs a focused audit. L3 — daily baseline at 03:00 scheduled via launchd runs audit-run.sh all. All three write one report per run to docs/audit/<date>-<focus>.md; CURRENT-ALERT.md is written only when the latest verdict is non-CLEAN, and the next interactive session is contractually obligated to read it before picking up the goal. — Three trigger layers, one output sink — README · Architecture / Design notes

Three-layer reactive audit

L1 — post-commit hook. Drift-risk commits (anything under agents/, .claude/agents/, config/, CLAUDE.md, the operator contract) fire audit-run.sh contract within ~30 s.
L2 — 15-min mission-anomaly poll. New blocker files or QA-FAIL bursts trigger a focused audit. No-op (zero tokens) when nothing's wrong.
L3 — daily 03:00 baseline. launchd fires the full sweep. Catches anything L1 + L2 missed.

The pattern is Reactor + Hook (files as events), not Observer — subagents in this repo aren't long-running observables.

Cost-routing rule

The architectural lesson from a real failure: applying "Tier 2 (local) = default" to every pipeline stage produces a quality ceiling.

Mechanical, high-volume stages (transcribe, render, fetch, beat-detect) — local. Token cost would be ruinous at scale.
One-shot creative stages (script hook, factual framing, mood-keyword extraction) — Claude. ~500 tokens per call, operationally negligible against the existing subscription quota, and quality compounds over the next 60 seconds of viewing.

Full reasoning in docs/cost-model.md.

— Skills & tooling · the portfolio

Skills portfolio: two production media skills, one Unity meta-skill in development, one parked experiment — each card shows what it takes in and ships out. music-video (Production): music file → 60-second 9:16 vertical short; 23 ffmpeg shaders, 7 grade profiles, 19 genre presets, cuts on aubiotrack beats, glitch on aubioonset drum hits, Shape A 5-agent missions pipeline. job-hunt (Parked): one seed keyword → deduped markdown digest; seed → 26-synonym role family, 11 source plugins, 5 live no-key / 2 key-gated / 4 mock, 5,000+ raw → ~200 matches, Shape B standalone curl+jq. game-dev-agent (In development): Unity meta-skill — sprite gen, C# scaffolding, balance tuning, audio gen, in-game AI director; validation surface = PawnSim colony-sim, 13-agent game roster, CLI-scaffolded with zero manual Unity Editor work. product-cf (Parked): one product photo → CF-style 9:16 short; parked on an honest negative finding — free/local 3D didn't clear the quality bar on a 16GB machine, needs paid cloud image-to-video or a bigger GPU. — Skills portfolio — what each skill takes in and ships out — README · Overview

music-video ✓ Production

Music file → 60-second 9:16 vertical short

23 ffmpeg shaders
7 grade profiles
19 genre presets
cuts on aubiotrack beats
glitch on aubioonset drum hits
Shape A · 5-agent missions pipeline

Skill #1, missions-routed; verified via a 12/12 Hermes drop-in interop test. Spec: skills/music-video/

job-hunt Parked

One seed keyword → deduped markdown digest

seed → 26-synonym role family
11 source plugins
5 live no-key · 2 key-gated · 4 mock
5,000+ raw → ~200 matches (mock run)
Shape B · standalone curl+jq

Skill #2, standalone, v2 short-keyword UX — parked: the KR job boards it targeted (Saramin, Wanted, JobKorea…) now block scraping, so it runs on mock/dry-run data only; the global-* ATS/RemoteOK/Remotive/HN plugins may still return live remote postings. 70 tests; 4 enrichment utilities behind JH_*_LIVE flags. Spec: skills/job-hunt/

game-dev-agent In development

Unity prompt → CLI-scaffolded game

validation surface = PawnSim colony-sim
13-agent game roster
zero manual Unity Editor work
not in the 2 production skills

Skill #3, the meta-skill above — sprite gen, C# scaffolding, balance tuning, audio gen, in-game AI Director.

product-cf Parked

One product photo → CF-style 9:16 short

parked on an honest negative finding
free/local 3D missed the bar on 16GB
needs paid cloud i2v or bigger GPU
kept in tree, gated off

Experimental v0.1.0 — realistic product motion needs paid cloud image-to-video the free/local path couldn't deliver.

Pipelines are packaged as portable Skills following the open agentskills.io standard — a skill written once can target multiple compatible runtimes (Claude Code, Cursor, Goose, Gemini CLI, OpenAI Codex, GitHub Copilot, etc.).

Skill #2 — job-hunt, in depth

A separate-shape skill: standalone (no missions-routed pipeline), v2 short-keyword UX, agentskills.io-compliant. Parked — the KR job boards it targeted (사람인 / 잡코리아 / 원티드 / 프로그래머스) now block scraping, so the live digest never came together; it runs only on mock/dry-run data. Pass --seed "Problem Solver" and the orchestrator expands to a 26-synonym role family (Forward Deployed Engineer / Applied AI Engineer / Generalist / Founding Engineer / …). A mock run de-dups 5,000+ raw postings down to ~200 matches; the code stays in the tree.

11 source plugins (5 live-ready without a key, 2 key-gated, 4 mock-fallback) — all mock-fallback by default; live HTTP per-plugin behind JH_SOURCE_LIVE=1.
4 enrichment utilities (fit-score / cover-letter / company-research / interview-prep) — per-posting Claude calls behind JH_TOOL_LIVE=1.
70 tests — smoke + edge-case + JSON-schema validation. EN + KO walkthroughs.

Quickstart: skills/job-hunt/scripts/run.sh --seed "Problem Solver" --dry-run. Reference digest at docs/samples/job-hunt-digest-mock.md.

Alternate path — faceless-short (narration era)

The earlier faceless-short mission still lives in the tree. Topic prompt in, narrated 60-second short out: Sonnet drafts the hook + factual framing, Kokoro-ONNX (English) or macOS Yuna (Korean) synthesizes voice, whisper.cpp transcribes for caption timing, Pexels B-roll selected per-window from caption keywords, ffmpeg burns single-line captions. Preserved for topic-driven content; not the current production format.

Pipeline detail in the README Sample output section.

Operator tooling

Scripts that surface system state and absorb routine status-check prompts, complementing the slow Claude-driven auditor with fast Claude-free checks.

scripts/first-touch.sh — zero-account demo wizard. Single Y/n consent; checks prereqs, fetches demo cache, renders, opens the result.
scripts/doctor.sh — ~2 s repo health check. CLI tools, env keys, schedulers, audit alerts, git state, disk, per-skill activation. --quiet / --json modes.
scripts/audit-skill-drift.sh — the 13th audit rule. Verifies each skill's declared LIVE-flag manifest matches the actual gating in its scripts.
scripts/statusline.sh — Claude Code statusline. Renders doctor:⚠N · goal:N/M · audit⚠ alongside dir / branch / model / cost.
scripts/morning-brief.sh — single-command overnight digest. Combines doctor + audit + 7-day intervention trend + commit attribution + autonomous decisions + review-queue + blockers into ~30 readable lines.
outputs/review-queue/ + 3 scripts — batched taste-decision queue. Operator drains a contact-sheet markdown on their cadence, ~10× fewer intervention events.

A meta-skill, goal-lock, parses docs/goal.md and reports unchecked deliverable subgoals so long autonomous sessions can re-anchor. Full per-tool table: docs/operator-tooling.md.

Evidence · what it produces

Evidence — what the pipeline actually produces

Frames captured from rendered mp4s. Full mp4s live under records/missions/ (gitignored — each ~ 25–50 MB).

music-video noir-detective — 2026-05-22 batch, t = 30 s. Per-genre grade_profile (rnb_low_key) shapes the pink-magenta low-key look; phrase-aware shader stack on top.

Genre catalog at a glance — six of the seven grade profiles

Mid-climax frames from the 2026-05-20 → 2026-05-23 production batch. No cherry-picked B-roll — every clip came from the same unattended Pexels mood-keyword fetch the pipeline always runs. The visual identity is the grade_profile + shader stack, not the source footage. Seven grade profiles compiled to ffmpeg filter graphs via scripts/music-video-grade.sh (six shown below); 19 genre presets in skills/music-video/data/genre-presets.yaml.

Genre catalog: each per-genre grade_profile recolors the same generic Pexels stock into a genre-coded look before the shader layer — no cherry-picked B-roll. Six profiles shown: noir-detective (rnb_low_key, deep magenta low-key), rain-lofi (lofi_warm_grain, soft warm grain), arcade-synthwave (city_pop_neon, purple → cyan neon), coastline-summer (hollywood_teal_orange, teal → orange split), linen-minimal (kr_warm_pastel, warm peach pastel), smallhand-folk (kinetic lyric overlay, Korean lyric overlay). 19 genre presets, 7 grade profiles (the 7th is a neutral ungraded passthrough). Identity is the grade + shader stack. — Same stock in, genre-coded look out — README · Genre catalog

noir-detective: smoky bar interior, bearded man with pipe, pink-magenta low-key grade — **noir-detective**
rnb_low_key

rain-lofi: barista pouring espresso, soft pink warm grade — **rain-lofi**
lofi_warm_grain

arcade-synthwave: VHS cassette + retro VCR in purple city-pop neon grade — **arcade-synthwave**
city_pop_neon

coastline-summer: golden-hour beach water reflections, hollywood teal-orange grade — **coastline-summer**
hollywood_teal_orange

linen-minimal: bedroom bookstack and coffee mug, kr warm pastel grade — **linen-minimal**
kr_warm_pastel

smallhand-folk: cafe through window with Korean lyric overlay visible — **smallhand-folk**
+ kinetic lyric overlay

Skill #2 — what a job-hunt digest looks like

Single seed keyword in, deduplicated markdown digest out. The mock-fallback rendering below comes from docs/samples/job-hunt-digest-mock.md — exercised against all five default sources, 26-synonym problem-solver family expansion, 7/21 raw postings matched. Real digests land under records/jobs/<date>/digest.md (gitignored).

# Job-hunt digest — 2026-05-20

> Seed: Problem Solver → role family problem-solver
>            (26 synonym keywords expanded)
> Sources: _mock, kr-wanted, kr-programmers, kr-jobkorea, kr-saramin
> Total postings: 7 — 0 new since last digest

### _mock (3)
- Problem Solver (AI Agent) · MockRebeatLike
  지역: 서울 강남구 · 게시: 2026-05-20
  요약: 쇼핑 AI Agent 기획+개발+배포까지 직접 담당.  PMF 탐색 사이클 주도.
- Forward Deployed Engineer · MockFrontierAI
  지역: 원격 · 게시: 2026-05-20
  요약: Build AI agent solutions; framing problems → shipping LLM
        prototypes within weeks.
- Generalist · MockKRStartup
  지역: 서울 마포구 · 게시: 2026-05-19
  요약: PM+Engineer+Data Analyst 하이브리드.  Ship MVPs, iterate to PMF.

### kr-wanted (1) · kr-programmers (1) · kr-jobkorea (1) · kr-saramin (1)
…

Faceless-short gallery (historical — narration-era pipeline)

Frames from the earlier faceless-short trials, retained as visual evidence of the narration-era pipeline that preceded the music-video pivot.

Frame from faceless-hittites-v6: 9:16, historical battle reenactment B-roll, caption 'Turkey and Syria for five centuries,' single-line. — **Hittites EN**

Frame from faceless-hydrogen-v6: pasta macro shot, caption 'The human body's reliance' single-line. — **Hydrogen EN**

Frame from faceless-autotune music-trial: studio audio waveform display, caption 'like a secret safety net. But in 1998,' single-line. — **AutoTune EN**

Frame from faceless-hittites-ko-v6: aerial Hattusa archaeological site, Korean caption single-line. — **Hittites KO**

Faceless-era scorecard historical

Self-evaluation across five retention-mapping axes (Hook, Visual sync, Readability, Factual coherence, Production polish), assigned by Claude during the faceless-short iteration — preserved as the structured progress signal from the v4 → v5 → v6 sequence that preceded the music-video pivot. The music-video mission uses platform watch-time data instead of per-dimension scoring; per-video metrics live under docs/pilots/.

Operator-intervention trend (autonomy signal)

A multi-agent system that needs constant human steering hasn't actually replaced the work it was meant to. Two-panel honest signal updated daily 02:00 KST — Panel A from git log (commit attribution + leverage ratio), Panel B from local Claude Code session JSONLs (operator prompt count + active session minutes). See case study #8 for the 5 prioritized reduction levers acting on the trend.

Mission-outcome trend (quality signal) historical · render-era

Historical (render-era) chart, last refreshed 2026-05-25. Since 2026-06 the active development focus shifted to the PawnSim game track, whose work shows up as verification-gated commits (see the autonomy trend above) rather than render missions — so this chart is no longer the current signal and is kept as a record of the render era, the same way the scorecard above is framed. It reads the system's render evolution at a glance: the 2026-05-17 spike (8 → 33 missions/day) is the faceless-pilot batch; the post-pivot flat band is the music-video format at a sustainable 3–8 renders/day cadence. Every records/missions/<date>/<id>/qa-report.md was parsed for Verdict: PASS|FAIL and attempt N of M.

Quickstart · try it

Try it in ~60 seconds

Single-command guided wizard. No Pexels signup, no Suno round-trip, no .env edit:

git clone --depth 1 https://github.com/MelonS/MelonS-Agents.git
cd MelonS-Agents
./scripts/first-touch.sh        # checks prereqs, fetches cache, renders, opens result

The wizard checks prerequisites, fetches the demo cache (~30 s), renders a 60-second 9:16 short from bundled CC-BY Blender clips + Kevin MacLeod tracks (~100 s), and opens the result. Single Y/n; rest is automatic.

Manual demo path (same end state)

./scripts/bootstrap.sh                # verifies tools, fetches models
MUSIC_VIDEO_DEMO_MODE=1 \
  ./agents/missions/music-video/run.sh demo

Skill #2 — job-hunt short-keyword demo (~5 s, no network)

skills/job-hunt/scripts/run.sh --seed "Problem Solver" --dry-run
# digest.md printed on stdout; mock-fallback postings spanning multiple
# sources, all matched against the 26-synonym "problem-solver" family.

Live HTTP per source (5 plugins require no API key): JH_GLOBAL_ATS_LIVE=1 JH_GLOBAL_REMOTEOK_LIVE=1 JH_GLOBAL_REMOTIVE_LIVE=1 JH_GLOBAL_HN_LIVE=1 JH_WORKNET_LIVE=1.

Full Pexels + Suno path (mood-keyword catalog, custom tracks) documented as the advanced path in the README Quick start. macOS first; Linux compatible for the core pipeline (whisper.cpp + ollama + ffmpeg + aubio), macOS-only for launchd schedulers and Yuna TTS.