Skip to content
Dev Kordeš
← Posts

Lunch Break Skill

LLM skill for when you have more token limits than you know how to use

I recently got myself a Max 20 Claude Code subscription. For 1 month, to finish a project. I also have 20x separately at work

And honestly, it's been a lot of fun

Most of the times I have tokens to spare before before the 5hr limit resets

So, I created a "lunch-break" skill. It's essentially a skill that tells the agent to go wild with exploration. Spawn many subagents and look into the topic you ask it to, no edits, just exploration

Then write a short summary and more thorough explanation into an .md file

Skill

Here's the version for my side project - github devk-skills

It has a couple of lenses and 2 modes: normal and max. Lenses are a11y, competitive (analysis), copy (text quality), SEO, UI and what-next (product manager)

Modes are how many subagents it's encouraged to spawn. Max goes wild wild, normal goes a little less wild

Does it work?

Yes and no. It finds a lot of irrelevant stuff, but also finds some genuine issues and potential improvements

At work, I have a similar skill, but focused on performance, bugs and reliability. That one genuinely found issues. It also helps when I point it to a specific part of codebase which has been neglected for years and could use a second set of eyes

And here's the TLDR example from running UI on my private. Mostly genuine issues:

# Lunch Break: ui
Generated: 2026-04-24 03:08
Intensity: maximum
Subagents dispatched: 26 (scout: 4, deep dive: 14 wave-1 + 8 wave-2)
Project: Standups

## TL;DR
- **Empty / loading / error states are missing or wrong on the public surface.** Backend search errors render as silent blank pages. `assign_async` `:loading` and `:failed` slots are literally empty strings. Release Radar renders heading + whitespace on empty data.
- The **Neo ↔ daisyUI design-system seam is leaking**: `core_components.ex` flash uses daisyUI `toast`/`alert` and shows up on neo-brutalist public pages on subscribe/unsubscribe error paths.
- **Browse — the primary discovery page — has no thumbnails.** Every card on every other surface has art. Browse is plain text rows.
- **Search dropdown overflows on mobile** and silently disappears on zero results — no loading indicator, no state feedback, stale highlights persist after typing.
- **Conversion funnels have inverted visual weight on the riskiest action** — the destructive "Unsubscribe from all emails" is styled as the safe ghost option.

Memories

It often finds things that just genuinely aren't an issue. Rerunning it again would then find the same issue

The thing I found best, is to adjust the skill to check a LUNCH_BREAK_MEMORIES.md before starting exploration. I then write whatever findings are irrelevant into that .md, so that next run ignores those

Try

If you're in similar situation, with tokens to spare, just create your own lunch break skill with your own needs and try it out. It's fun :)