Crispy | AI Builder

Building apps with an AI dev team 🤖 Practical Claude Code & vibe-coding tricks you can actually copy.

[Crispy | AI Builder @blackputiii]

@ASH_MERCER_ Did not know this existed. A ramp-up guide generated from the actual CLAUDE.md and hooks sounds way more honest than the onboarding doc nobody updates.

[Crispy | AI Builder @blackputiii]

@RoundtableSpace The before/after score is the part I like. An agent saying 'I made it faster' means nothing, an agent showing the new Lighthouse run is something you can actually check.

[Crispy | AI Builder @blackputiii]

@thepanta82 This is the hardest kind to catch because it works. Tests pass, review looks fine, then six months later nobody understands why admin logic lives in the shared server.

[Crispy | AI Builder @blackputiii]

@whoizmide Painfully accurate. I've started naming the files it is allowed to touch before anything else. 'Only edit X, touch nothing else' saves so many surprise diffs.

[Crispy | AI Builder @blackputiii]

@0xJoseDinero The delete test is brutal but right. If the model already does it by default, the rule is just context tax. I'd keep a tiny safety net and rebuild the rest from actual mistakes.

[Crispy | AI Builder @blackputiii]

Stop prompting Claude Code like a search box. Give it a finish line: 'Done when the test passes and you show me the diff.' Half the bad output I see is just a missing definition of done.

[Crispy | AI Builder @blackputiii]

Credit to Lydia Hallie from the Claude Code team: https://t.co/CRIhpZuP0m

[Lydia Hallie ✨ @lydiahallie]

We recently added idle compaction to Claude Code!

It now compacts long idle sessions right before the prompt cache expires (~1h). If you come back to it later, it only has to reprocess the summary instead of the entire conversation, which saves you some usage.

[Crispy | AI Builder @blackputiii]

Left a fat Claude Code session open through lunch?

Used to hurt when you came back. Prompt cache dies after about an hour, so your next message reprocessed the whole chat at full price.

In 2.1.299 long idle sessions (~200K+) now compact first, roughly 54 minutes in, right before the cache expires. You resume from the summary instead of the full history.

Catch: summaries drop details. Put the stuff that must survive (paths, rules, decisions) in CLAUDE.md, not only in the chat.

[Crispy | AI Builder @blackputiii]

@levelsio That's a second reason to own the boring internal tools. Curious if you've seen founders actually turn down those offers, or if the cash usually wins.

[Crispy | AI Builder @blackputiii]

@ClaudeDevs The "see what each session is doing" bit is what I wanted from the CLI. Do nested sessions share CLAUDE.md / memory, or does each one start clean?

[Crispy | AI Builder @blackputiii]

@kentcdodds The same-branch case is the interesting one. Do the agents know how to rebase, or does Kody pause them and let one land first?

[Crispy | AI Builder @blackputiii]

@tdinh_me In-context is the part that always broke for me with the SaaS tools. How are you feeding the surrounding UI copy into the harness so "Save" doesn't become "Speichern" in a button that should stay English?

[Crispy | AI Builder @blackputiii]

@tristanbob Same logic applies to the sidebar tab, it keys chats to the folder that's open. Open each project as its own window via Open Folder on the root, not one shared multi-root workspace, and the history should split per project.

[Crispy | AI Builder @blackputiii]

@lydiahallie The ~1h timing is the smart part. It lines up with the prompt cache expiring, so coming back costs you a summary instead of 200K tokens. Pairs well with ending sessions on a small cleanup skill.

[Crispy | AI Builder @blackputiii]

Stop retyping your best prompt.
Turn it into one file.

Save it as .claude/skills/techdebt/SKILL.md. From then on you just type /techdebt.

Example: at the end of a session it finds duplicated or dead code in today's changes, suggests the smallest fix first, and changes nothing until you say go.

The idea is from the Claude Code team's own tips. Credit below.

[Crispy | AI Builder @blackputiii]

Theo's new benchmark has Gemini 4 Argon getting more than half of its claims about a codebase wrong.

That isn't only a Gemini problem. Any agent can sound sure about code it never opened.

The rule I'd add to every codebase question:

"For every claim, cite file:line. If you didn't find it in the code, say 'not found'."

Then open two of the lines it cites. If they don't match, stop trusting the answer.

[Crispy | AI Builder @blackputiii]

@theo How does it handle a long agentic session in a real repo? Does it stay consistent after 30+ tool calls, or does it start drifting like earlier Gemini models?

[Crispy | AI Builder @blackputiii]

@theo A >50% false claim rate on codebase analysis is the number that matters for agent work. Great at one-shot demos means nothing if it confidently describes code that isn't there. Did it get better when you made it cite file paths?

[Crispy | AI Builder @blackputiii]

@gregisenberg Same product, new customer, new story. Indie devs skip this constantly. We build for "everyone" when one narrow audience would actually pay.