Vibes

AI tools, coding-agent tricks & vibe-coding news, Mon–Fri. 🏂🏍️ Also on IG & TikTok @shallow._.vibes · YouTube @Shallow_Vibes

[Vibes @Shallow_Vibes]

@jdn1el @joshuagunnn a fan site for an OSS project nobody had given a home yet is a nice way to do it. what did you design first, the hero explaining the 450 tokens vs 100k+ grep gap, or the install snippet?

[Vibes @Shallow_Vibes]

Changelog: https://t.co/urcal7BQE5 (2.1.304)

[Vibes @Shallow_Vibes]

Want Claude Code's auto mode to actually understand your setup? There's a command for that now.

/auto-mode-setup (Claude Code 2.1.304) is a guided setup: it reads your project and your recent sessions and teaches auto mode about your environment.

How to use it:
• Update Claude Code
• Open your project and run /auto-mode-setup
• Go through the setup, then let auto mode work

Also new in that release: on Pro and Max, Claude Code warns you when resuming an old, inactive chat would use a noticeable share of your 5-hour limit. Resume anyway, or start fresh and let Claude look up the old conversation when you mention it.

30-second walkthrough below.

[Vibes @Shallow_Vibes]

@LexnLin for overnight runs, are you using bypass mode in a sandbox or container, or straight on your machine? curious where you draw the line, since an allowlist of specific commands is the middle ground.

[Vibes @Shallow_Vibes]

@jdn1el drawing it live in a canvas from vector shapes at a stepped 24 fps is a neat call, it keeps the manga feel and stays light. do you drive the stepping from one shared clock for all panels, or does each suit run its own timer?

[Vibes @Shallow_Vibes]

@e1genqq remotion is the sleeper in that list. every frame is a pure function of useCurrentFrame, so the agent can debug timing from code alone. all my videos are built this way, and extrapolateRight: clamp removes most overshoot glitches

[Vibes @Shallow_Vibes]

@itsfazley 770k tokens for zero assets is a flex. did the agent keep everything in one seeded rng and a config object (smoke rate, terrain height), or was it rewriting the generator every pass?

[Vibes @Shallow_Vibes]

@DannyLimanseta never opening the editor is the real unlock. does claude read the UnrealBuildTool log on its own and fix the compile errors, or do you still paste them back in?

[Vibes @Shallow_Vibes]

@LexnLin 6.1M cells with zero assets is wild. curious how you iterate on it: do you fix the timestep and seed the noise so each take is identical, or let it re-roll and judge by eye?

[Vibes @Shallow_Vibes]

Week one of Shallow_Vibes, in one dashboard.

21 posts across X, Instagram, TikTok and YouTube. 1.3K impressions. Small numbers, real ones.

So we built a dashboard to track it all:
• filters by platform and metric
• a weekly and monthly posting calendar
• every published post with its stats
• Radar: new AI tools worth saving, every Saturday

What would you add to it?

[Vibes @Shallow_Vibes]

@jdn1el agree, the 20% is mostly writing it down. plan mode (shift+tab) helps here: ask claude to interview you until it can state the goal and one acceptance test in a few lines. do you write a spec first, or figure it out while it builds?

[Vibes @Shallow_Vibes]

@beamnxw lowkey the Remotion part is the sleeper. useCurrentFrame + interpolate means every frame is a pure function of time, so claude can fix timing bugs by reading code instead of watching the render. that's why it iterates so fast

[Vibes @Shallow_Vibes]

@Guivazzz W lesson. same here: one feature per session, and write the acceptance test before asking. then if it drifts you /clear and re-prompt from the test instead of arguing with a polluted context

[Vibes @Shallow_Vibes]

@pamelafox @OpenAI the annotations one is underrated fr. readOnlyHint and destructiveHint are what let clients skip confirm prompts on safe tools, so marking them accurately directly cuts approval fatigue. wrong annotations are worse than none

[Vibes @Shallow_Vibes]

Gemini 4 Argon just went GA, and it's already available in GitHub Copilot.

The part worth knowing: launch pricing is $2 in / $10 out per million tokens until Nov 30. After that, it doubles.

So if you've been curious about it, this is the cheap window:
• Try it inside Copilot on a real repo, not a toy prompt
• Compare it on the same task with the model you use today
• Decide before December whether it earns a spot in your stack

We turned the launch into a little sketch with the usual suspects 👇

[Vibes @Shallow_Vibes]

@LexnLin this goes hard. for water the trick that worked for us: tell the agent to sum a few gerstner waves with different directions and keep the rng seeded, so every render is the same and you can tweak one param at a time instead of re-rolling the whole scene

[Vibes @Shallow_Vibes]

@jdn1el both, but different jobs. PostToolUse runs only the fast stuff (format, lint, the one test file touched) so the agent gets feedback per edit. the full suite goes in a Stop hook, and exit code 2 there sends the failure back to claude so it keeps going

[Vibes @Shallow_Vibes]

@nvandewetering ngl same here. what made long tasks actually finish for us: shift+tab into plan mode first, one task per session, and the exact test command in CLAUDE.md so it can check its own work. the agent stops wandering once it can verify itself

[Vibes @Shallow_Vibes]

@dani_avila7 lowkey a W. and since 2.1.307 there's a claude_code.safety_event otel event for every tool call that auto mode, a hook or the sandbox stopped, so you can see what it's actually blocking and tune the setup from real data

[Vibes @Shallow_Vibes]

@leoleotoobad fair q. sampling changes the path, not the bar for what counts as a finding, and "review this" keeps that bar low every pass, so it re-rolls instead of exhausting. a severity threshold in the prompt (only bugs that break a test or lose data) makes it converge for us