Deep-Decode is an opinionated, end-to-end AI content factory.

Give it a blog post, a tweet, a product, a person, or a trending topic. It researches the source, writes an opinion-dense deep-dive (not a translation, not a summary), draws comic-style infographics, lays out an illustrated document, voices a podcast, cuts a video, and pushes every format to publish-ready across email, WeChat, Xiaohongshu, Video Accounts and Douyin.

The whole run is orchestrated by a declarative skill graph and a deterministic runner that refuses to advance until each step's artifact passes its contract. No silently-skipped steps. File = state. Failure is visible.

This is not a demo. The output/ folder holds 130+ finished pieces shipped through the pipeline.

Every figure below is generated by the GPT-Image (img2) backend — accurate Chinese typography, a consistent palette, magazine-grade layout. These are real, recent pieces (June 2026), straight from the pipeline. Click any image to read the decode.

“The referee can't be the player” Why an agent needs a harness, not a nicer model.	“The bottleneck moved” When writing code is free, what becomes the constraint.	Ideogram v4.0 Text-to-image learns to spell and follow structure.
Fei-Fei Li · world models A concept taxonomy, drawn as one circular diagram.	Anthropic Partner Hub Tiered staircase + headline metrics.	Zhipu Z.AI ⟷ Anthropic Dual-column comparison, two IPO paths.

Each deep run emits a full img2 figure set (cover + per-section illustrations, like the hero grid above), plus a podcast (podcast.mp3), a video (video.mp4), an illustrated Word doc (.docx), and magazine-style Xiaohongshu card decks — all addressed by the same output/<slug>/ contract. The backend is switchable per project via image_backend: gpt-image.

The old pipeline was ~18 prose steps in a doc. Agents running a deep chain would silently get lost — skip a step, use the wrong template, stop at an email draft instead of shipping. So the flow became a declarative graph (skillgraph.yaml) traversed by a deterministic runner (tools/pipeline.py).

            ┌──────────────────────────────────────────────────────────┐
            │  skillgraph.yaml  —  declarative 3-layer graph            │
            │                                                          │
   compounds│  decode · brief · practice · distribute-all   ← you drive │
            │      ▲  (pick a playbook + confirm Strategy Spec)         │
  molecules │  article · infographic set · podcast · video · doc        │
            │      ▲  (each = one deliverable)                          │
     atoms  │  svg→png · tts · imagegen · tone-lint · send              │
            │      ▲  (single deterministic action)                     │
            └──────────────────────────────────────────────────────────┘
                         │  topological order computed by
                         ▼
            tools/pipeline.py  —  the ONLY authority on step order
                 status · next · gate · verify

• Three layers. atoms (one deterministic action) → molecules (one deliverable) → compounds (a full playbook a human drives).
• You don't count steps. The runner computes the topological order from depends_on edges.
• Done = artifact exists and passes its contract. Contracts include file_exists, min_bytes, png_for_each_svg, audio_visual_sync, tone_match (banned-word + voice lint), and channel_draft_ready. A missing artifact = stuck; the runner won't continue.
• Humans steer at the compound layer only — choose the playbook, confirm the Strategy Spec. Everything else the runner pulls along.

cd output/2026-06-08-some-slug
python3 ../../tools/pipeline.py status   # whole graph + ✓/✗ + next step
python3 ../../tools/pipeline.py next      # "what node runs now?"
python3 ../../tools/pipeline.py gate m.article   # verify the artifact contract

A few of the 130+ pieces (titles translated; sources are decoded in Chinese):

Browse them all in output/, or through the web workbench below.

A Next.js 15 app (web/) turns the factory into a three-in-one site:

Static-generated, zero media in the deploy bundle — images/audio are rewritten to a jsDelivr CDN backed by this repo, video to GitHub raw. Deploys to Vercel with deploy as the production branch.

cd web
npm install
cp .env.example .env.local
npm run dev          # http://localhost:3000

deep-decode/
├── .claude/skills/   # the skills: deep-decode, polish, visual, podcast, video, distribute…
├── skillgraph.yaml   # the declarative 3-layer graph (single source of step order)
├── tools/            # pipeline.py runner + atoms (tone_lint, tts_atom, imagegen_relay…)
├── output/           # 130+ finished pieces — one folder each (article + media)
├── wiki/             # structured knowledge: topics / sources / concepts / published
├── schedule/         # queue · in-progress · published · calendar
├── styles/           # voice + feedback + best-of, multiple kits
├── readers/          # audience personas (drives tone)
├── templates/        # content + project templates
└── web/              # Next.js 15 portal + process view + admin

Deep-Decode runs as a Claude Code skill system.

1. Clone the repo and open it in Claude Code — the skills live under .claude/skills/.
2. Drop source material in raw/, or hand it a URL.
3. Say "decode this" (/deep-decode <url>) and confirm the Strategy Spec; the runner takes it from there to publish-ready.
4. Run the web workbench (cd web && npm run dev) to browse what you've made.

Heads-up: distribution targets Chinese platforms (WeChat / Xiaohongshu / Video Accounts / Douyin) and Chinese TTS voices. The architecture — declarative skill graph, contract gates, one-source-to-many-formats — is fully general; fork the pattern for any language or channel set.

Actively used in production for a Chinese AI-commentary publication. deploy is the content branch (the default branch on GitHub); see web/README.md for why content lives on an orphan branch and how it's pushed.

Apache-2.0 © 2026 AmoryMing. The pipeline, skills, and tooling are free to use and adapt. The decoded articles and generated media under output/ are illustrative examples — please don't republish them as-is.