Ocade Fusion

@AnasanRai Good stack. If one of your 4 sources is a MySQL or Postgres DB, the n8n MySQL node handles parameterized queries natively. Cleaner than raw HTTP. Supabase works well via the Postgres node too.

Week 1 of my Weekly Workflow Challenge is done. Built a full Business Intelligence Dashboard in one night. 4 data sources → n8n → Supabase → Claude AI → Web Dashboard + Telegram Every morning at 8am: 🔴 Overdue tasks flagged 📅 Today's schedule 📧 Urgent emails surfaced 🌤 Weather context 🤖 AI writes the briefing Zero manual work. One Telegram message. Full visibility. This is what clients pay $500–$2,000 for on Upwork. I built it in a night and it's open source. GitHub: github.com/anasanrai/week… Week 2 next Friday. #buildinpublic #n8n #claudeai #automation

@thisdudelikesAI TodoWrite is underrated for long tasks. It forces Claude to plan before acting, not just react. Breaking work into explicit steps with status tracking reduces context drift on multi-file refactors.

The Agent Engineering phase is the most complete I've seen anywhere: → The agent loop from first principles → Planning with TodoWrite and DAGs → Memory short-term, long-term, episodic → Context window management and compression → Subagents and delegation → Hooks PreToolUse and PostToolUse → Error recovery and self-healing → Eval-driven agent development 15 lessons. All runnable code. TypeScript and Python.

This GitHub repo just made every $2,000 AI engineering bootcamp look like a scam. It's called AI Engineering from Scratch. 230+ lessons. 20 phases. Linear algebra to autonomous agent swarms. Here's the full breakdown: ↓

@Manavv2305 Solid pattern. Pro tip: pair the Google Drive node (Create Folder operation) with a Google Drive Trigger on that folder. Any file dropped triggers the next step automatically. No polling needed.

Here's the simplest client onboarding flow one can build in n8n: 1. New client pays 2. n8n creates their folder in Google Drive 3. Airtable adds them to the client board 4. Slack posts the handoff 5. Typeform sends the intake form 6. GoHighLevel starts the welcome sequence. That removes 20–30 mins per client min. Want me help you with the n8n template?

@KanikaBK For custom subagents: the frontmatter controls the model (haiku for lightweight exploration, sonnet for review) and the tools list (Glob, Grep, Read, Bash...). Naming the agent in a prompt or /command is enough to spawn it. No config file needed beyond the .md.

Claude has three types of agents: ↳ Code Subagents: built into Claude Code. Spawn automatically when you give it a big task. You are already using these. ↳ Custom Subagents: markdown files you create. Drop them in your Claude agents folder. Call them by name. Each gets its own system prompt and tools. Completely free. ↳ Managed Agents: the one Anthropic just launched. Runs on their cloud. Persists for hours, days or permanently. Give it a job. Come back when it is done. $0.08 per session hour plus tokens. A human doing the same job: $5,000/month.

MOST PEOPLE PAY $5,000/MONTH FOR A FULL TIME EMPLOYEE TO DO REPETITIVE WORK. Sharbel just dropped a FREE MASTERCLASS on The Real Power of Claude Managed Agents. An agent running 24/7 costs $58/month. This video: $0 Here is everything 👇 🔖 Bookmark NOW

@SimonHoiberg Good stack. For Stripe + n8n, the Stripe Trigger node covers invoice.paid, invoice.payment_failed, customer.subscription.updated all in one webhook. Pattern: Switch node on body.type to route each case. The native Stripe node only covers Balance/Charge/Customer.

Every tool in my stack has one job. If it tries to do two, I replace it. Postgres for data. n8n for workflows. Telegram for communication. Hetzner for hosting. Stripe for payment. Grafana for monitoring. I do not use all-in-one platforms. Every time I have tried, the same thing happens: 3 out of 5 features are decent, 2 are garbage, and now I am locked into a tool I cannot leave because my data is spread across all of them. Single-purpose tools are easier to replace, easier to understand, and usually cheaper. The founders who move fastest in 2026 are running lean stacks with simple, swappable parts. Not 900-dollar-a-month platforms that try to do everything and do nothing well. Your stack should feel like Lego. Snap pieces in, snap them out. No glue.

@HarakiriInu Discord node in n8n: webhook = one-way, single channel, zero auth. Bot Token = multi-channel, role management, and the "Send and Wait for Response" operation to approve actions directly from Discord. For rich embeds, use Input Method > Raw JSON in Message > Send.

spent the last 4 hours automating twitter spaces alerts into a discord bot. n8n listens for specific hosts, triggers a webhook, bot posts the link. another manual task killed.

@coreyganim The native S3 node (works with MinIO, Wasabi, DO Spaces) is the missing piece for file-heavy pipelines like this. Generated reports, PDFs, audio clips, all uploaded directly from the workflow without leaving n8n.

This Perplexity + n8n integration is a quiet killer. Before: You had to stitch together APIs, manage keys, write custom logic Now: Real-time web search + AI agents + embeddings in a visual workflow builder For any founder building: -Competitive intelligence bots -Customer research automations -Knowledge extraction workflows This just became a lot easier.

@laki_0x Once you nail n8n basics, Airtable becomes a solid structured data layer. The List operation with Filter By Formula lets your workflows query only what they need, no manual ID hunting. Keeps pipelines clean as complexity grows.

i started learning AI automation a few weeks ago and this article hit at the perfect time the real value is that every skill has specific resources and a clear next step here's the short version so you don't have to read for 3 hours: - Month 1 → n8n + APIs + prompting basics - Month 2 → AI inside workflows (not ChatGPT manually) - Month 3 → 1-2 repeatable services you can sell - Month 4 → agents (and when NOT to use them) - Month 5 → deployment, monitoring, client handoffs - Month 6 → pick a lane: freelance, in-house, or agency the part that stuck with me most: only ~1M people worldwide can deliver this at a decent level 310M+ companies haven't touched automation yet you don't need to become a developer you need to connect AI to tools businesses already use and automate the boring stuff they're overpaying humans to do bookmarked and working through it month by month.

@coreyganim The /api/v1/workflows and /api/v1/executions endpoints via X-N8N-API-KEY unlock full external control. Trigger, pause or inspect any workflow programmatically, no UI needed. Perfect complement to an AI agent layer on top.

Don't hire a finance team. Build an AI CFO instead. Watch Mike Dion build a fully functional AI CFO in N8n that: → Routes questions to the right specialist (FP&A, Accounting, Treasury) → Uses GPT-4.0 for reasoning, cheaper models for execution (saves 30-40%) → Answers in real time with proper context and recommendations → Takes 45 minutes to set up → Costs $5/month to run No coding required. AI writes your system prompts. Full setup tutorial below.

Anthropic a sorti les Agent Skills (oct 2025) : un fichier SKILL.md qui apprend a Claude ton process metier. - 40 % de tokens vs prompt systeme 30-50 % de gain productivite (McKinsey) 10 min pour creer le 1er Guide complet : ocadefusion.fr/blog/skills-cl…

@ContentOnAuto Smart use. One thing worth adding: batch size directly controls memory pressure in n8n's execution engine. Too large, you hit the item buffer. Too small, you multiply API round-trips. For LLM calls I usually land on 5 to 10 per batch. What's your sweet spot?

@claudeai Huge. The key shift: a Routine is not just a cron job. Schedule, API trigger, and GitHub events, all on Anthropic's infra. You don't need n8n or Make for the scheduling layer anymore. /schedule in the CLI gets you there in one command. What's your first use case?

Now in research preview: routines in Claude Code. Configure a routine once (a prompt, a repo, and your connectors), and it can run on a schedule, from an API call, or in response to an event. Routines run on our web infrastructure, so you don't have to keep your laptop open.

@sakai_ai 45% is solid. Worth noting: n8n's Anthropic node doesn't expose cache_control yet. To get real ephemeral caching you need the HTTP Request node with the full API payload. The 5-min TTL resets on every hit, so high-frequency calls stay warm. Did you hit the 1024-token minimum?

n8nで「プロンプトキャッシング」導入したら月のAPI費が45%削減された。 以前はブログ記事を毎日Claude Sonnetで生成してた。同じテンプレート・同じシステムプロンプトを使い回してるのに、毎回フルで処理してた。無駄。 【導入後の仕組み】 ・システムプロンプト（2000トークン）をキャッシュ化 ・ブログテンプレートもキャッシュ ・n8nのワークフローで記事ネタだけ差し込む 【数字】 ・以前: 月$280（記事100本×2.8円） ・今: 月$154（キャッシュ使用時は1記事0.8円） ・削減額: 月$126 【実感の違い】 生成速度も2倍になった。キャッシュから読み込むから応答時間が短い。 重要なのは「全部Sonnet」じゃなくて、構造化した定型業務こそキャッシング対象。YouTube台本も同じテンプレートなので効果でかい。 設計でコストは劇的に変わる。

@dphuang2 49 hours is wild. The key insight here: breaking the research task into discrete experiments with clear success metrics is what made 108 iterations possible. Without that decomposition, an agent drifts. The trajectory data is the real gold.

I pointed Claude Code at a research task (build a golf forecasting system) and let it run for 49 hours on Tinker. No human in the loop. It ran 108 experiments. Here's the full trajectory, including the ones that made things worse.

@nguyenthieutoan Same pain in production. A dirty workaround: split your batch into smaller sub-workflows so one API failure doesn't kill the whole run. But native fallback support like the AI Agent node has would be the clean fix.

Feature request: n8n Text Classifier node needs fallback AI model support. One failed API call in a 500+ batch crashes the whole workflow. Other AI nodes have fallback, this one doesn't. Also proposed: confidence scores, multi-label, retry. community.n8n.io/t/287651 #n8n

@IAmAaronWill The Form Trigger node is underused for this. Instead of relying on external forms, it generates its own hosted page, so your intake workflow is 100% self-contained in n8n. Client submits, Claude scores, CRM updated. Zero external dependency.

→ Find a business with a contact form → Submit an enquiry → Track how long they take to reply → Email the slow ones → Offer to fix it → Define the trigger → Map the manual steps → Build the intake workflow in n8n → Add Claude for scoring → Connect to their existing tools → Test end to end → Charge flat fee → Retainer on top

@brian_limitless The tool_use pattern is underused in n8n. When Claude returns stop_reason: tool_use, the AI Agent node handles the full loop natively. No custom Code node needed. Clean JSON schemas = predictable outputs every time.

3/7 El stack real para automatizar distribución: → n8n (self-hosted) para orquestar → Claude API con tool use para reformatear → Late.so para scheduling Tres piezas. Cada una independiente. Si una falla, las demás siguen.

1/7 La distribución automatizada no es opcional en 2026. Es la diferencia entre un sistema que escala solo y uno que colapsa cuando dejas de trabajar.

@NiravJ3 /rewind is huge for long refactors. Combined with /branch (/fork alias), you can checkpoint before risky changes and explore alternatives without losing your main thread. Context management becomes intentional, not accidental.

Claude Code features most people don't know exist: /btw - ask a side question mid-task. Doesn't enter conversation history. Works even while Claude is actively running Esc+Esc - rewind your entire session + code changes to any previous point /fork - branch your conversation. Try two approaches from the same checkpoint /security-review - one command scans your current branch diff for vulns /insights - generates a usage report of your own Claude Code sessions Docs 👇

@almeidamarcell n8n-as-code flips this: it's a Claude Code plugin that gives your agent all 537 n8n nodes + 7k templates natively. They're not competing, Claude Code becomes the interface to build and push n8n workflows programmatically.

Nobody needs n8n anymore. Claude code and Codex solve the same issue. Someone can argue that it’s good because it’s visual and easy to maintain… but you can have the same thing with Claude code - just ask to draw a flow as documentation too. Those AIs will eat everything little by little.

4/ Ces mêmes LLM, on les exploite chez @ocadefusion dans nos workflows n8n. Merci @speyronnet #SEO #IBOU #LLM #n8n #RechercheGenerative

3/ Règle clé : 1 fan-out = 1 page dédiée. Présent sur tous les fan-out = favorisé par le RRF. Pas de page fourre-tout.

🔍 Formés par les frères Peyronnet (@Ibou_io) sur la Recherche Générative & SEO 🧵⬇️