Pier Paolo | AI & Vibe Coding

Dal consulente al builder AI. Vibe coding, prompt system, marketing che vende.

[Pier Paolo | AI & Vibe Coding @PieMedici]

Dal consulente al builder AI.

Pubblico: vibe coding, prompt system, marketing che vende — errori e vincoli inclusi.
Non pubblico: liste di tool, motivazionale, news digerite.

Se fai lo stesso passaggio, dimmi a che punto sei bloccato — rispondo sotto.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@rzrgrv Permissions first for me too. On SchedaFit the guard looked fine, then nav still opened a screen wider than the token. Shared role map for auth + nav fixed it — more prompts didn’t.

[Pier Paolo | AI & Vibe Coding @PieMedici]

Grok + Cursor → SchedaFit (Android gym plans). UI fast.

Broke on master/trainer/client: login OK, nav used a different scope than the token.

Fix: one shared role map for auth + nav — not another prompt.

Multi-role: did screens or permissions fail first?

[Pier Paolo | AI & Vibe Coding @PieMedici]

@carolinaorsinyf Su SchedaFit il prompt system ha spedito le schermate. Si è rotto sui ruoli master/trainer/client: auth e nav non condividevano lo stesso scope. Lì un altro prompt non bastava — una sola role map sì.

[Pier Paolo | AI & Vibe Coding @PieMedici]

Dal consulente al builder AI.

Pubblico: vibe coding, prompt system, marketing che vende — errori e vincoli inclusi.
Non pubblico: liste di tool, motivazionale, news digerite.

Se fai lo stesso passaggio, dimmi a che punto sei bloccato — rispondo sotto.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@mustafarazameer AST blast-radius before context is the adult version of “paste the repo.” On a multi-role app the wrong files aren’t just waste — they teach the agent the wrong permission surface.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@hailports Green loop vs done work — same trap as agent “booking confirmed.” I gate ship on owner-visible outcomes, not agent status. If the back-office isn’t filed, the dashboard is fiction.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@maxinwild This is the only BIP that scales for me: draft from the Cursor trail, edit once, ship. Manual “build in public” always loses to the next auth bug. Byproduct > content calendar.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@rang_manthan Living this on Android: login/JWT fine, then master vs trainer vs client broke nav. Token had a role; the graph didn’t. Your JWT→ROLE_ mapping is the missing half — without it the UI lies about what you can do.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@stretchcloud @shadcn Same failure mode as multi-role auth: green login ≠ correct action. My agent “succeeded” when nav opened a screen the token shouldn’t allow. Fix wasn’t a better prompt — it was refusals + role checks before writes.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@frrntc Hit this on a gym app: UI came fast, then master/trainer/client broke because auth and nav assumed different scopes. “Vibe” got screens; it didn’t get a shared role map. Quality showed up the second permissions mattered.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@SystemArch_AI Ran a similar audit after stacking skills on a Grok+Cursor Android build. Dead skills were “free”; the ones that fired still ate context every turn. Rule I keep: if a skill doesn’t cut a decision or a refusal, unload it.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@dteam69 @luciascarlet Got it — so the practical line is: Agent SDK harnesses OK, OMP-style use is “not allowed but tolerated.” I’ll stick to one Anthropic entity either way.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@codewithkhalil Appreciate it — the hard part is forcing the agent to list refusals before it writes. Without that, “trade-offs” are just vibes.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@dteam69 @luciascarlet Good to know — so the billing split risk was more “early chaos” than a lasting rule. I’ll keep one Anthropic entity anyway.

[Pier Paolo | AI & Vibe Coding @PieMedici]

Grok + Cursor → SchedaFit, app Android (presto anche su IOS) per schede palestra.

Schermate veloci. Blocco: ruoli master/trainer/client (pezzi ok, flusso no).

Regola: zero nuove schermate finché i 3 ruoli non condividono lo stesso grafo auth.

Tu: rotto su UI o su permessi?

[Pier Paolo | AI & Vibe Coding @PieMedici]

@rohanpaul_ai @bcherny Deleting the file isn’t the point — seeing which constraints were load-bearing is. Keep the 3 rules that change output; drop the rest as costume.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@neembu_paani31 @cybersaksham The failure mode isn’t bad code — it’s approving a plan you didn’t read. Gate: no agent write until the plan has 1 constraint you would defend in a review. Without that, “architecture first” is just another prompt.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@FReza1984 Prompt monolitico invecchia in 2 settimane. Folder skill sì: versionabile, ispezionabile, rimovibile. Test: se non puoi diffare la skill come codice, non è un sistema — è un memo.

[Pier Paolo | AI & Vibe Coding @PieMedici]

@GohilHardy Shipping faster isn’t the bug — zero ownership after merge is. Rule that stuck: every vibe PR names a maintainer + a kill date for the shortcut. No name = no merge.