convez.ai
Agente
Cheeryoh

launch-content-studio

Cheeryoh/launch-content-studio

Ver no GitHub
InicialPluginScore · de 0 a 100

HITL multi-agent orchestration: product-launch brief to coordinated marketing & education asset package. Runs on a Claude Code Max plan.

Instalação

$npx convez-ai Cheeryoh/launch-content-studio

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Launch Content Studio

A human-in-the-loop, multi-agent orchestration that turns a product-launch brief into a coordinated marketing & education asset package — in one supervised run.

Built for the workflow a Customer Education Program Manager owns at launch: take a fast-moving product update and produce the use cases, positioning, blog post, social copy, announcement email, webinar abstract, and demo script that customer-facing teams need — without losing the human quality bar.

Part of a three-product portfolio. See also Drift Sentinel (product↔course drift) and Signal Loop (customer signals → education priorities).


Why it exists

OpenAI ships fast. The bottleneck isn't writing one asset — it's producing a consistent, accurate, multi-channel package quickly, every time, while keeping a person accountable for what goes out. This pipeline does the drafting and the consistency checking; a human approves (or rejects with feedback) at four gates. Reject once and the responsible agent re-runs with your feedback as context — so the human stays in control without doing the keystrokes.

The pipeline

 BriefParser ───────────────────────────────▶ [Gate 1 · Brief]
 AudienceSegmenter → Positioning ───────────▶ [Gate 2 · Positioning]
 AssetWriter ───────────────────────────────▶ [Gate 3 · Asset drafts]
 MessagingQA ───────────────────────────────▶ [Gate 4 · Final sign-off]
                                                     │
                                                     ▼
                                           coordinated asset package
AgentRole
BriefParserAgentRaw launch notes → structured LaunchBrief (what changed, why it matters, proof points, audience, channels)
AudienceSegmenterAgentBrief → enterprise segments by jobs-to-be-done, with one designated primary
PositioningAgentBrief + audience → positioning statement, value pillars (each tied to a proof point), reusable key messages
AssetWriterAgentProduces the full package: blog, social, email, webinar abstract, demo script
MessagingQAAgentIndependent skeptical reviewer — scores each asset on clarity/accuracy/audience-fit and flags unsupported claims

Each gate's payload is shown to the reviewer in the dashboard. Reject → the agent re-runs with your reason (and optional stated expertise, which is weighted into the rewrite). This is the anti-sycophancy loop: the agent is explicitly told to produce a materially different result, not to repeat itself.

Run it

Requires Node 20+. No API key needed — by default every agent call runs through your local claude CLI (Max plan), so there's zero API billing.

npm install
npm run dev          # http://localhost:3000

Open the dashboard, paste a launch description (or click an example), and approve/reject each gate.

LLM backends

The backend is auto-selected; override with LLM_BACKEND:

BackendWhenCost
cli (default)local claude CLI presentnone (Max plan)
apiANTHROPIC_API_KEY setAPI usage
fixtureUSE_FIXTURES=truenone — canned outputs for offline/demo/test

Test

npm test            # full pipeline, fixture mode, auto-approves every gate
npm run test:reject # same, but rejects the positioning gate once to exercise the rewind path

Both assert that all five artifacts are produced and the run completes.

Architecture notes

  • Promise-based HITL gates (src/kernel/HITLGate.ts): the conductor awaits a gate promise that only resolves when a human calls the approve/reject route. The pipeline genuinely cannot advance without a decision.
  • Dual-backend client (src/kernel/ClaudeClient.ts): CLI / API / fixture behind one interface.
  • Zero-dependency persistence (src/kernel/storage.ts): each run's state, artifacts, and an append-only audit log live under .data/<pipelineId>/.
  • Live progress over Server-Sent Events; the dashboard reconnects mid-pause and replays the open gate.
  • The kernel (src/kernel/) is shared verbatim across all three portfolio products; only agents.ts, schemas.ts, Conductor.ts, and public/config.js differ.

Project layout

src/
  kernel/        ClaudeClient, BaseAgent, HITLGate, ConductorBase, SSEManager, storage, server, types
  agents.ts      the 5 product agents
  schemas.ts     Zod artifact schemas
  Conductor.ts   the launch pipeline (stage + gate sequence)
  index.ts       wires the conductor into the kernel server
public/          dashboard (index.html, app.js, style.css, config.js)
fixtures/        canned agent outputs for offline tests
scripts/         testPipeline.ts (end-to-end fixture test + rewind)

MIT licensed.

Ver o original no GitHub

Bom para

Relacionados