DEV Community: Hong Phuc

How We Built a 5-Step Cross-Platform Publishing Pipeline with AI Agents

Hong Phuc — Mon, 08 Jun 2026 07:30:42 +0000

Most developer blogs start with a markdown file and end with a "Publish" button. Ours starts with an AI agent scraping GitHub Trending and ends with the same post live on four platforms — each with the correct canonical URL, proper SEO ownership, and a Supabase record tying it all together.

Here's how we built it, what broke, and what we learned.

The Problem

We publish a GitHub Trending digest three times a week. Each post needs to live on:

hongphuc5497.com — our personal site (Next.js, Vercel)
Substack — newsletter subscribers
Dev.to — developer community
Supabase — content store for the site's dynamic pages

Doing this manually meant: copy-paste to Substack, format for Dev.to, write a YAML-frontmatter note for the site, then forget to update the content database. Canonical URLs were wrong. Dates drifted. The Supabase store was perpetually out of sync.

We wanted a single command that produces all four — in the right order, with the right URLs.

The Architecture

The pipeline has five steps, executed in strict order:

Step 1: Create   → LLM agent writes the post (Hermes cron)
Step 2: Site     → hongphuc5497.com (canonical source of truth)
Step 3: Substack → canonical URL → personal site
Step 4: Dev.to   → canonical URL → personal site
Step 5: Supabase → records all 3 platform URLs

The personal site is always first. Substack and Dev.to both point their canonical_url to it. This means Google indexes the personal site — not the platforms. SEO ownership stays with us.

What We Automated

Three AI agents handle different parts:

Hermes runs the cron scheduler — fetches trending data, orchestrates the pipeline, delivers to Telegram
Codex builds the repo infrastructure — staging scripts, archive logic, Agent Ops protocol
The Hermes agent itself composes the digest analysis — reads the raw JSON, writes 2-3 sentence summaries per repo

The fetcher script scrapes github.com/trending directly (no API key needed), extracts repo names, star counts, languages, and descriptions, then outputs structured JSON with local timezone fields. The agent uses those fields — not UTC, not the system clock — for the post title and filename.

What Broke

1. The Date Drift Problem

The most subtle bug: the trending data was fetched at 10:36 PM on June 4, but the cron job ran at 9:00 AM on June 5. The agent titled the digest "June 05" even though the fetcher data said "June 04."

The fix: Two-layer enforcement. The staging script now exits with code 1 when the title date doesn't match the fetcher's local_date_long field. The cron prompt explicitly warns: "STAGING WILL FAIL with exit code 1 if the title date does not match the fetcher date."

This turned a silent warning into a hard pipeline failure. The agent can't ignore it anymore.

2. The Supabase Namespace Shadow

The markdown-content-store repo has a supabase/migrations/ directory (standard Supabase CLI layout). Python's import system treated this as a namespace package, which shadowed the real supabase pip library.

When any publishing script ran from supabase import Client, Python found the empty migrations directory instead of the installed library. Every content store save silently failed.

The fix: In src/db.py, we prepend .venv/lib/python*/site-packages to sys.path before the import. The real library is found first. One file, 14 lines, zero external changes.

3. The Multi-Line List Numbering

The Hermes agent writes numbered lists like:

1. ***chopratejas/headroom*** · Python · 12.6K⭐ +3.1K Compress tool outputs before they hit the LLM...

But when Substack's from_markdown() parser sees this, it merges the continuation line into the heading, breaking the numbering entirely. The same issue hit Dev.to.

The fix: We built a custom ProseMirror JSON builder (pm_builder.py) that correctly handles headings, inline formatting, and multi-line content. For Dev.to, we added _reformat_ordered_lists() to merge continuation lines into single-line items before hitting the API.

The Canonical URL Chain

hongphuc5497.com/notes/{slug}     ← SOURCE OF TRUTH
        │
        ├── Substack              ← canonical → personal site
        └── Dev.to                ← canonical → personal site

Substack and Dev.to both declare the personal site as their canonical source. Search engines index the personal site. The platforms get the content but defer SEO authority.

The Supabase content store records all three URLs via a published_urls table — one post, multiple platform entries. The site reads from Supabase to populate its notes page dynamically.

The State Machine

trending.json → digest.md → stage → archive → deliver
                    │
                    ├── run.json      (date_consistent, warnings)
                    └── publish.json  (status: draft → delivered → published)

The staging script is the gatekeeper. It checks date consistency, copies artifacts into the canonical state directory, records metadata in run.json, and — if --auto-archive is set — archives the digest to docs/ with an updated index.json.

If anything fails, the pipeline stops. No partial publishes. No stale state.

What We Learned

Python namespace packages are silent killers. A supabase/ directory with no __init__.py will shadow your pip install with zero error messages. Always check import supabase; print(supabase.__path__) when imports break mysteriously.
Prompt enforcement is not enough. We told the agent "use ONLY the fetcher's date" and it still inferred from the system clock. Hard exits are the only reliable enforcement.
Canonical URLs must be set at publish time, not after. If Substack publishes without a canonical URL, you can't retroactively add one without deleting and republishing.
The personal site must be first. If you publish to Substack first, its URL becomes the de facto canonical — and you lose SEO ownership to your own platform.
Content stores are boring infrastructure that saves you hours. The Supabase upsert_by_slug + add_platform_url pattern means re-runs don't create duplicates, and the site always knows where every post lives.

The Result

Three times a week, an AI agent scrapes GitHub Trending, writes a 10-repo digest, and publishes it to four platforms in under 2 minutes. The personal site owns the canonical URL. The content store tracks everything. And if the dates don't match, the pipeline refuses to proceed.

No manual copy-paste. No stale Supabase records. No wrong canonical URLs.

Just an agent, a fetcher, and a very opinionated staging script.

The publishing pipeline is open source in github-digest. The content store lives in markdown-content-store. Both repos use Hermes Agent for orchestration.

📊 GitHub Trending — June 08, 2026

Hong Phuc — Mon, 08 Jun 2026 06:39:51 +0000

A curated look at the hottest open-source repos. No fluff.

mvanhorn/last30days-skill* · Python · 32K⭐ +1.1K An AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web — then synthesizes a grounded summary. The "last 30 days" framing forces recency, making it a research time machine rather than another static knowledge base. Surging interest signals demand for agents that can actually *look things up instead of hallucinating from training data.
Leonxlnx/taste-skill* · Shell · 37.3K⭐ +1.1K Gives your AI agent *taste — stops it from generating boring, generic slop. Written in pure Shell, it's a personality injection layer that enforces style constraints before the LLM speaks. The explosive growth shows developers are done tolerating bland AI output and want agents with actual editorial voice.
RyanCodrai/turbovec · Python · 7.6K⭐ +1.6K A vector index built on TurboQuant, written in Rust with Python bindings. The fastest-growing repo today — its Rust core + Python bindings pattern mirrors what made Polars and Ruff successful. Signals that the vector search space is maturing past "good enough" into raw performance competition.
NousResearch/hermes-agent · Python · 186.4K⭐ +1.1K The agent that grows with you — self-extending AI assistant framework with skills, memory, multi-agent orchestration, and cron scheduling. This is the engine behind the very digest you're reading. Consistent 1K+ daily growth reflects real adoption, not hype.
opencv/opencv · C++ · 88.2K⭐ +65 The 12-year-old Open Source Computer Vision Library is trending again — likely driven by new edge-AI use cases and the convergence of LLMs with visual processing. A reminder that foundational infrastructure never stops being relevant.
aaif-goose/goose · Rust · 47.7K⭐ +322 An open-source, extensible AI agent that goes beyond code suggestions — install, execute, edit, and test with any LLM. Built in Rust for speed. The "any LLM" design avoids lock-in while the execution-first approach makes it a serious alternative to cloud-locked coding assistants.
lfnovo/open-notebook · TypeScript · 27.5K⭐ +554 Open-source reimplementation of Notebook LM with more flexibility and features. Lets you build your own research assistant that synthesizes across documents with source-grounded responses. Growing fast as developers realize Notebook LM's closed nature limits what they can build on top.
Crosstalk-Solutions/project-nomad · TypeScript · 29.9K⭐ +309 Project N.O.M.A.D — a self-contained, offline survival computer with critical tools, knowledge, and AI. Offline-first, self-contained, and designed for when the internet isn't there. The "nomad" framing resonates with the growing desire for AI that works without cloud dependency.
yikart/AiToEarn · TypeScript · 19K⭐ +183 Let's use AI to Earn — an automation platform for monetizing AI workflows. The blunt naming says everything about where the AI community's head is at right now: tooling → revenue, not tooling → demos.
ggml-org/llama.cpp · C++ · 115.5K⭐ +158 LLM inference in C/C++ — the backbone of local-first AI. Steady growth reflects the ongoing shift toward on-device and edge inference. Not explosive, but foundational — every local agent stack depends on it.

📊 GitHub Trending — June 05, 2026

Hong Phuc — Mon, 08 Jun 2026 04:02:49 +0000

chopratejas/headroom · Python · 12.6K⭐ +3.1K — Compress tool outputs, logs, files, and RAG chunks before they hit the LLM — 60-95% fewer tokens with identical answers. Ships as a library, proxy, and MCP server, making it pluggable into virtually any LLM pipeline.
NousResearch/hermes-agent · Python · 181K⭐ +1.9K — The agent that grows with you — Hermes continues its meteoric rise as the go-to local agent platform. Skills, memory, multi-agent orchestration, and cron scheduling make it a complete alternative to cloud-locked coding assistants.
affaan-m/ECC · JavaScript · 207K⭐ +1.8K — An agent harness performance optimization system bundling skills, instincts, memory, security, and research-first development. Works across Claude Code, Codex, Opencode, Cursor, and beyond.
PaddlePaddle/PaddleOCR · Python · 79.9K⭐ +141 — Battle-tested OCR toolkit that turns any PDF or image into structured data for AI consumption. Supports 100+ languages and bridges the stubborn gap between scanned documents and LLM-ready text.
github/spec-kit · Python · 109K⭐ +321 — GitHub's official toolkit for Spec-Driven Development — define specs, generate implementation plans, and keep AI coding agents aligned with intent rather than guessing.
NVIDIA/cosmos · Jupyter Notebook · 9K⭐ +133 — NVIDIA's open platform of world models, datasets, and tools for building Physical AI. Targets robots, autonomous vehicles, and smart infrastructure with pretrained models.
lfnovo/open-notebook · TypeScript · 25.1K⭐ +212 — Open-source reimplementation of Notebook LM with more flexibility and features. Lets you build your own research assistant that synthesizes across documents with source-grounded responses.
Open-LLM-VTuber/Open-LLM-VTuber · Python · 9.6K⭐ +581 — Talk to any LLM hands-free with voice interaction, interruption, and Live2D avatars — all running locally across platforms. Brings the VTuber aesthetic to open-source voice AI.
jwasham/coding-interview-university · 350K⭐ +632 — The legendary computer science study plan continues to attract engineers preparing for technical interviews. A complete, self-paced curriculum for breaking into big tech.
github/copilot-sdk · Java · 9K⭐ +38 — Multi-platform SDK for embedding GitHub Copilot Agent into apps and services. Signals GitHub's push to make Copilot an embeddable platform rather than just an editor plugin.

AI Tools & Products Radar — May 28, 2026

Hong Phuc — Thu, 28 May 2026 17:40:52 +0000

A weekly snapshot of new AI tools, products, and platform launches that matter for builders.

This week in one sentence: AI coding tools command billion-dollar valuations, agentic AI moves from demos to enterprise, and Google's AI search is driving users to DuckDuckGo.

I've been tracking AI product launches through Firecrawl and TechCrunch feeds over the past week. The signal is clear: the AI race has shifted from models to distribution. Here's what caught my attention.

🤖 AI Coding Tools Are Still the Hottest Category

Cognition raised $1 billion at a $25 billion pre-money valuation. That's a coding assistant startup valued higher than many public SaaS companies. OpenRouter doubled its valuation to $1.3 billion in just one year — turns out model routing and API gateway infrastructure is a real business.

Meanwhile, Microsoft shipped Copilot Cowork — an autonomous multi-step AI agent built directly into Microsoft 365, in collaboration with Anthropic using Claude technology. It can execute complex workflows across Office apps without hand-holding.

Figma Make now edits production codebases. You can visually tweak a UI and it modifies the actual source code, not just mockups. Google AI Studio launched "vibe coding" mode with a Google AI subscription, and Colab Learn Mode shipped with Gemma 4 — an open model that is byte-for-byte the most capable open model right now.

🧠 AI Agents Go Mainstream

Google's Cloud Next 2026 was all about agents. The Gemini Enterprise Agent Platform lets organizations build, deploy, and manage fleets of AI agents. Google also rolled out their 8th-gen TPUs — custom silicon purpose-built for the agentic era, not just traditional ML training.

Meta launched "Plus" tier subscriptions across Instagram, Facebook, and WhatsApp, with AI plan trials bundled in. The social giant is betting you'll pay for AI features alongside ad-free browsing.

Robinhood now lets AI agents trade stocks autonomously. Users can delegate trading decisions to bots. The risks are obvious — but so is the signal: every platform is becoming an AI agent platform.

On the acquisitions front, Publicis bought LiveRamp for $2.2 billion specifically to strengthen its position in "agentic AI business systems." The ad industry sees AI agents as the next wave of marketing automation.

🎨 Creative AI Gets Legit

ElevenLabs Music v2 can switch genres mid-song — opera to heavy metal in the same track. It's trained on licensed data and cleared for commercial use. Not a toy anymore.

Gemini Omni, Google's anything-to-anything multimodal model, deepfaked a reporter in front of the Eiffel Tower during a live demo. Impressive and unsettling in equal measure. Meanwhile, ChatGPT now lives in PowerPoint — a sidebar that generates and edits presentations from prompts.

Amazon Prime Video greenlit three AI-made animated series through their GenAI Creators' Fund. The line between "AI-assisted" and "AI-generated" is dissolving fast.

🔬 Hardware Arms Race

The semiconductor boom continues. Snowflake signed a $6 billion deal with AWS for AI CPU chips. Micron and SK Hynix both crossed the $1 trillion market cap threshold, driven entirely by AI memory demand.

ByteDance is developing custom CPU chips for their AI rollout (Reuters exclusive). Vertu — yes, the luxury phone brand — launched the Alphafold, a $6,880 AI foldable with a built-in "AI Hermes" assistant meant for CEOs.

🔍 Search Is the New AI Battleground

The most interesting tension right now is in search. Google went all-in on AI Overviews at I/O, and users are pushing back. DuckDuckGo reported a 30% spike in iOS installs and a 27% jump in "No AI" search visits immediately after Google's AI announcements.

Google responded by adding Preferred Sources — users can now pin trusted sites in AI Overviews. Click-through rates reportedly double when a source is marked as preferred. CNN, meanwhile, is suing Perplexity over "verbatim" scraping of paywalled content. Copyright in the AI era hasn't been solved — it's being litigated one lawsuit at a time.