Conversation vs specification: the real line between vibe coding and agentic coding
'Vibe coding is a conversation. Agentic coding is a specification.' Everything else about the transition follows from that one distinction.
Vibe coding: you explain a requirement informally, the AI builds, every new feature is another chat turn — 'you don't have clarity; you're trying to get there.' The platforms named for it: Bolt, Lovable, Emergent. Agentic coding: you SPECIFY — 'in your head it is absolutely clear what you want to build. You queue the specification, and sometimes even a swarm of agents builds your app.' The kitchen version: grandma's recipe ('this much... then this much... feel it, you know when it's done') is a conversation — always decent, never replicable; the MasterChef recipe ('150 degrees, 12 minutes — not 10, not 15') is a specification — perfectly replicable, and fragile: 'any one step goes here and there, your dish is off.' Neither is 'better'; they're methods — but 'if you're trying to become an AI generalist, start making the transition.'
The honest nuance: 'the lines can be blurred — you could be vibe coding in Cursor and attempting agentic coding in Lovable.' The transition has three steps: get comfortable with GitHub, learn to write good specs, then learn agentic orchestration (Windsurf/Devin). And the fragility of specs is exactly why the four concerns (edge cases, auth, rate limits, error handling) exist — specification precision cuts both ways.
His own household: 'my wife is very specific; I do wild cooking. We made a pact never to be in the kitchen at the same time' — the two methods, married.
It gives the course's scattered vibe/agentic vocabulary one clean test: is this a conversation or a spec?
Agentic coding is vibe coding with a better model.
It's a different input artifact: a specification with edge cases, auth, limits, and error behavior — the model matters less than the clarity.
Vibe coding is a conversation. Agentic coding is a specification.
Grandma's recipe is very good — but replication is going to be tricky. Specification is replicable. And any one step goes here and there, your dish is off.
Your batched Lovable briefs are specifications by his definition — this session is the theory under your lovable-cost-discipline batching rule.
Go deeper
In one line: Vibe coding = iterative conversation toward an unclear goal (Bolt/Lovable tier); agentic coding = queued specification executed by agents (Cursor/Windsurf/Claude Code tier); specs replicate but are fragile, so spec-writing skill (edge cases, auth, limits, errors) is the actual transition.
The share-your-app test opens the session: 'has your friend come back saying this is not working?' — the gap this sprint exists to close ()
Spec fragility is stated as a feature-cost: precision buys replication at the price of brittleness ()
Anxiety permission granted twice: 'these are optional sessions... drop off anytime' — the sprint-tier teaching contract ()
▶ Watch this taught:
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
You're three prompts into 'make it nicer.' Which mode are you in, and what would switching cost?
Conversation. Switching means stopping to write what DONE looks like — features, constraints, error behavior — which feels slower and is cheaper by the third feature.











