Day 8: AI Workflows — Chaining Tools, the IDEA(L) Analysis Framework, Claude Skills vs MCP vs Projects, and the Reel-Cloning Pipeline

'The power of chaining': the output of one AI tool becomes the input of the next, which 'overcomes single tool limitations.' The loop: break the task into sub-steps, find the best tool for each, connect them, refine on results. The worked example: research in Perplexity or Gemini Deep Research -> outline in Claude or ChatGPT -> visuals in DALL-E or Nano Banana Pro -> layout in Canva. Dileep ties it back to the MCP session's stacking principle and warns it is learned by experimentation - 'don't get frustrated early.' The mindset: orchestrator of specialists.

Identify and Import: state the goal precisely and have the model produce a 5-8 question 'learning agenda' - 'this entire analysis hedges on how you are setting the goal.' Describe and Discover: profile columns, types, quality, and list interesting questions the data could answer. Explore and Examine: patterns, hypotheses, and explicitly 'questions we cannot answer' from this data (benefits, time-to-hire, turnover). Analyze and Apply: visuals and summaries that pass the '5-second rule' - 'can someone understand this chart immediately?' Learn and Leverage: 100% human - domain expertise, storytelling, context; 'the human sets the goals... the human is the judge.'

Live in Claude on a Kaggle data-science salaries CSV reframed as HR-to-CEO budgeting: the generic dataset produced a specific finding (fully remote workers earning ~16% more), hypotheses (a blended team cuts cost 30-40%; the 2022 45% spike signals a talent war), budget scenarios, and a CEO narrative. The technique he calls visual prompting: attach the screenshot of each framework step and instruct 'follow only this, don't analyse yet.' Alternate split: ChatGPT/Gemini for the first three steps, Julius or SuperJoin for the last two.

Claude is the chef - it thinks, decides, creates. A Skill is a recipe card: step-by-step instructions for a repeatable task. MCP is kitchen equipment: the capability to reach external systems. A Project (or custom GPT) is the station: a workspace bundling skills and tools for one ongoing context (a client, a campaign, quarterly planning). A Skill has three components: metadata (name + description, ALWAYS loaded so Claude can match your prompt and invoke it automatically), instructions (the 'expert prompt'), resources (templates, examples, scripts). Decision rule: same instructions repeated across contexts -> Skill; need an external system -> MCP; ongoing specific context -> Project.

Live: Settings > Capabilities > Skills > example skills > Skill Creator > 'try in chat' produced a 'Resume Tailor' skill that turned a vanilla CV + job description into a tailored resume. Skills are paid-tier; the free workaround copies skill.md from Anthropic's public skills GitHub repo into ChatGPT (strongest/thinking model) as a meta-prompt - losing auto-invocation. To 'a glorified prompt creator' Dileep says yes: it is prompt engineering, packaged, reusable, and auto-triggered.

Five stages as run by Vaibhav's team. IDEA: monitor news, filter for what you actually find interesting (not virality), add the audience angle (here 'an Indian perspective'); running example, a Sergey Brin AI story. SCRIPT: reverse-engineer hooks from high-view YouTube tutorials by pulling their transcripts (Google AI Studio + Gemini Flash, or a transcript extractor) into a reusable hook prompt; generate hook + body; then rewrite into the creator's 'writing DNA' from five past Reel transcripts. 'The voice is the non-scalable part.' The first five seconds decide retention - hooks get ~80% of scripting effort. FOOTAGE: 'first, create the voice' in ElevenLabs - IVC (~2 min sample) vs PVC (~2 hours, used for the main voice); they stayed on multilingual v2 because v3 alpha 'sounded artificial'; polish the TTS by italicizing/capitalizing words in a doc and removing em-dashes; regenerate until right. Then HeyGen: keep several monthly 'looks' (front-facing, laptop angle); upload the finished AUDIO so lips sync to the real take; lighting and angle decide fidelity. EFFECTS: Higgsfield for VFX from start/end frames (slow); Suno Pro or ElevenLabs Music for tracks. EDIT: Premiere (paid), DaVinci Resolve (free), Veed.io for subtitles/eye-contact/green-screen; CapCut is banned in India. Editing 'was not very, very effective' with AI - 'the human in the loop becomes very, very important.'

Definitions as given: an API is two-way request/response used when your app must actively request or send data (an agent calling the Gmail API); a webhook is a one-way, event-triggered push into your workflow (n8n firing on a new email or form fill) - 'synchronous vs asynchronous'; MCP is the protocol an LLM/agent uses to call external tools. Sub-workflows and sub-agents are locked to one platform; the same capability as an MCP server is portable across workflows and platforms. Amandeep's example: the Explorium MCP keeps prospect data fresh and flags funding or job-change events so outreach reads 'congrats on your funding.' Why not one universal MCP? Overloading a single agent with everything confuses it - 'if you bombard the LLM with so many MCPs... it's not gonna give us optimum results.' In the main session MCP is re-stated as 'a layer between your AI agent node in n8n and the tools' that stops tool-list hallucination.

Hallucination: unclear frameworks, no output format, no chain-of-thought steps, and long chats degrading recall - restart the conversation. CLAUDE.md files persist context across sessions. Build your own MCP server from the official Model Context Protocol docs, or hand the folder to Cursor/Antigravity and prompt it to deploy - 'no coding required'; a WhatsApp MCP server was installed live by prompting Cursor with a public repo link. Tool picks at recording: Lovable, Bolt, Emergent, Same.dev for apps (with Claude Opus 4.5 / Codex 5.2 as the coding models); Meshy AI 'number 1' for 3D; Groq for free daily LLM credits vs OpenRouter paid; Bolna (YC) and Retell for Indian voice numbers; Unipile for unofficial personal WhatsApp automation versus the regulated Business API; Apollo API over cookie-based LinkedIn scrapers; Office 365 Copilot has no exposed API - use Claude/OpenAI or Ollama/GPT-OSS locally. Error handling in n8n: a dedicated error workflow per client routed to Slack/email. IP on AI-made products: file if valuable, laws vary, 'distribution matters more than IP protection.'