Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Plan and run multi-agent video production pipelines.
.claude/skills/nousresearch-kanban-video-orchestrator/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 11% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 310% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 118% | 0% |
Wrap any video request — from a 15-second product teaser to a 5-minute narrative short to a music video to an ASCII loop — in a Hermes Kanban pipeline that decomposes the work to specialized agent profiles.
This skill does not render anything itself. It is a meta-pipeline that:
The actual rendering happens inside the kanban once it's running, via whichever existing skills + tools fit the scenes — ascii-video, manim-video, p5js, comfyui, touchdesigner-mcp, songwriting-and-ai-music, heartmula, external APIs, or plain Python with PIL + ffmpeg.
ascii-art, gifs, meme-generation, songwriting-and-ai-music).ascii-video).DISCOVER → BRIEF → TEAM DESIGN → SETUP → EXECUTE → MONITORThe discovery process is adaptive: ask only what is actually needed. Always start with three questions to identify the broad shape:
From the answer, classify the style category. The style determines which follow-up questions to ask. Do not ask all questions at once. Ask 2-4 at a time, listen, then proceed. Make reasonable assumptions whenever the user implies an answer.
For complete intake patterns and per-style question banks, see references/intake.md.
Once enough is known, produce a structured brief.md using the template in assets/brief.md.tmpl. Stages:
Show the brief to the user for confirmation before designing the team. The brief is the contract — every downstream task references it.
Pick role archetypes from the library that fit this video. Compose, don't clone. Most videos need 4-7 profiles. The director is always present; the rest are picked by what the brief actually requires.
For the role library and per-style team compositions, see references/role-archetypes.md.
For mapping role → which Hermes skills + toolsets it loads, see references/tool-matrix.md.
Generate a setup script (setup.sh) and run it. The script:
~/projects/video-pipeline/<slug>/)taste/, audio/, assets/hermes profile create --cloneSOUL.md (personality + role definition)brief.md, TEAM.md, and taste/ contenthermes kanban create task assigned to the directorUse scripts/bootstrap_pipeline.py to generate setup.sh from a brief + team-design JSON. See references/kanban-setup.md for the setup script structure, profile config patterns, and the critical "shared workspace" rule.
Run setup.sh. Then provide the user with monitoring commands:
bashhermes kanban watch --tenant <project-tenant> # live events hermes kanban list --tenant <project-tenant> # board snapshot hermes dashboard # visual board UI
The director profile takes over from here, decomposing the work and routing tasks to specialist profiles via the kanban toolset.
Stay engaged — the kanban runs autonomously but a stuck task or bad output needs human (or AI) judgment.
Monitoring patterns: poll kanban list periodically, inspect any RUNNING task that exceeds its expected duration with kanban show <id>, and check heartbeats. When a worker's output fails review, the standard interventions are:
kanban_comment)For diagnostic patterns, intervention recipes, and the "task is stuck" playbook, see references/monitoring.md.
Six concrete pipelines covering very different video styles — narrative film, product/marketing, music video, math/algorithm explainer, ASCII video, real-time installation — showing how the same workflow yields very different teams and task graphs. See references/examples.md.
asking at least the three baseline questions. A bad brief cascades through the entire pipeline.
every job. A music video that doesn't have a beat-analysis profile will misfire. A narrative film that doesn't have a writer profile will produce incoherent scenes. See references/role-archetypes.md.
dir: workspace. Tasks pass artifacts via shared filesystem and structured handoffs. Every kanban_create call passes workspace_kind="dir" + workspace_path="<absolute project path>".
(--tenant <project-slug>). Keeps the dashboard scoped and prevents cross-pollination with other ongoing kanbans.
relevant renderer should load that skill via --skill <name> on its task or always_load in its profile. Do not re-derive what a skill already provides.
file toolset, the director's SOUL.md rules forbid it from executing work itself. It decomposes and routes only — every concrete task becomes a hermes kanban create call to a specialist profile. The kanban orchestration guidance auto-injected into every kanban worker's system prompt spells this out further.
Aim for the smallest task graph that still parallelizes well and exposes the right human-review gates.
image-to-video) need keys in ${HERMES_HOME:-~/.hermes}/.env or the user's secret store. A worker that hits a missing-key error wastes a task slot. The setup script's check_key helper aborts cleanly if a required key is missing.
SKILL.md ← this file (workflow + rules)
references/
intake.md ← discovery question banks per style
role-archetypes.md ← role library (writer, designer, animator, …)
tool-matrix.md ← skill + toolset mapping per role
kanban-setup.md ← setup script structure & profile config
monitoring.md ← watch + intervene patterns
examples.md ← six worked pipelines
assets/
brief.md.tmpl ← brief skeleton
setup.sh.tmpl ← setup script skeleton
soul.md.tmpl ← profile personality skeleton
scripts/
bootstrap_pipeline.py ← generate setup.sh from brief + team JSON
monitor.py ← polling + intervention helpers| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 18,532 | 5,559 | -70% | 1 | 1 | 0% | 3,452 | 3,274 | -5% | 0 | 0 | — |
case-02 | fail→pass | 10,387 | 6,058 | -42% | 1 | 1 | 0% | 1,965 | 3,314 | +69% | 0 | 0 | — |
case-03 | fail→pass | 16,668 | 6,483 | -61% | 1 | 1 | 0% | 3,023 | 3,341 | +11% | 0 | 0 | — |
case-04 | pass→pass | 7,827 | 3,865 | -51% | 1 | 1 | 0% | 1,488 | 2,949 | +98% | 0 | 0 | — |
case-05 | fail→pass | 27,019 | 9,116 | -66% | 1 | 1 | 0% | 972 | 3,984 | +310% | 0 | 0 | — |
case-06 | pass→pass | 9,097 | 3,410 | -63% | 1 | 1 | 0% | 1,586 | 2,774 | +75% | 0 | 0 | — |
case-07 | fail→pass | 6,308 | 1,806 | -71% | 1 | 1 | 0% | 1,158 | 2,521 | +118% | 0 | 0 | — |
case-08 | pass→pass | 5,911 | 1,890 | -68% | 1 | 1 | 0% | 958 | 2,581 | +169% | 0 | 0 | — |
case-09 | fail→pass | 7,735 | 2,330 | -70% | 1 | 1 | 0% | 1,535 | 2,607 | +70% | 0 | 0 | — |
case-10 | fail→pass | 7,472 | 6,127 | -18% | 1 | 1 | 0% | 1,393 | 3,352 | +141% | 0 | 0 | — |
case-11 | pass→pass | 14,208 | 11,175 | -21% | 1 | 1 | 0% | 2,531 | 4,350 | +72% | 0 | 0 | — |
case-12 | fail→pass | 9,740 | 1,468 | -85% | 1 | 1 | 0% | 1,532 | 2,467 | +61% | 0 | 0 | — |
case-13 | fail→pass | 3,890 | 2,666 | -31% | 1 | 1 | 0% | 708 | 2,710 | +283% | 0 | 0 | — |
case-14 | fail→pass | 9,592 | 1,885 | -80% | 1 | 1 | 0% | 1,571 | 2,526 | +61% | 0 | 0 | — |
case-15 | fail→pass | 5,694 | 2,262 | -60% | 1 | 1 | 0% | 969 | 2,639 | +172% | 0 | 0 | — |
case-16 | fail→pass | 9,349 | 2,422 | -74% | 1 | 1 | 0% | 1,593 | 2,680 | +68% | 0 | 0 | — |
case-17 | pass→pass | 7,264 | 5,386 | -26% | 1 | 1 | 0% | 1,158 | 3,029 | +162% | 0 | 0 | — |
case-18 | fail→pass | 6,883 | 1,723 | -75% | 1 | 1 | 0% | 1,208 | 2,506 | +107% | 0 | 0 | — |
case-19 | fail→pass | 7,028 | 3,243 | -54% | 1 | 1 | 0% | 1,252 | 2,848 | +127% | 0 | 0 | — |
case-20 | fail→pass | 10,237 | 1,475 | -86% | 1 | 1 | 0% | 1,866 | 2,483 | +33% | 0 | 0 | — |
case-21 | pass→pass | 11,506 | 6,534 | -43% | 1 | 1 | 0% | 1,984 | 3,450 | +74% | 0 | 0 | — |
case-22 | fail→pass | 5,655 | 1,924 | -66% | 1 | 1 | 0% | 1,052 | 2,575 | +145% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +73 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/8/2026 | +64% |
Other measured skills in the registry, with their headline benchmark lift.