Ash Explained

I test AI agents, prompts, and coding workflows so builders know what actually works.

[Ash Explained @AshExplained]

@jackfriks @postbridge_ Jack do you think giving 50% off first month or more if they let you know why they left and join back again is a good idea to reduce churn?

[Ash Explained @AshExplained]

@rileybrown I feel claude cowork and GPT work should make themselves as how grok bot did. GPT even has agents in chatgpt web which they should reskin similar to grokbot

[Ash Explained @AshExplained]

@trq212 @bcherny I would love a similar product like grok bot by Anthropic...any plans to do such things?

[Ash Explained @AshExplained]

@mattpocockuk Ah yes that makes sense. Haven't thought about that lol

[Ash Explained @AshExplained]

@mattpocockuk can you please update prototype skill to have an option where it lays down UI screens and their connections to other UIs in an infinite canvas like figma does?

[Ash Explained @AshExplained]

I have moved from terminal to desktop app of claude code and it has its own perks..loving it so far especially the browser preview and artifact

[Ash Explained @AshExplained]

Anyone else uses claude skills to file GST?

[Ash Explained @AshExplained]

@aaliya_va Exactly. AI is useful for getting unstuck, not for outsourcing your taste.

The final draft still has to sound like a person wrote it.

[Ash Explained @AshExplained]

Peak adhd is creating systems with ai agent before even deciding what's the end product will look like. I should first record some youtube videos and edit manually before creating systems to do it for me

[Ash Explained @AshExplained]

Since Roblox now supports native MCP I created a skillpack for complete #Roblox game development and maintenance. Enjoy!

https://t.co/832W1WGEjC

[Ash Explained @AshExplained]

The June 25 Codex release made one workflow lesson click for me:

An agent should not need a giant helper list before it can start.

Better pattern:
You describe the job.
It looks up the right helper when needed.
The main context stays cleaner.

That matters because a crowded prompt makes small mistakes more likely.
It also makes tool swaps less painful later.

If your setup still depends on remembering exact tool names every run, the workflow needs another pass.

What helper do you still have to remind your agent about every single time?

[Ash Explained @AshExplained]

@sflorimm Tiny internal tools keep winning because they remove one annoying repeat task fast. Polished demos get attention, but boring painkillers tend to stick.

[Ash Explained @AshExplained]

@Amank1412 The first useful signal usually comes before the first 1,000 users. A few real replies from people with the pain teach more than broad posting numbers.

[Ash Explained @AshExplained]

@codingnninja I would not throw prompt and context work away that fast. Tiny changes there decide whether the agent opens the right files or spends 20 minutes confidently being wrong.

[Ash Explained @AshExplained]

@paraschopra Self-only tools are a great fit for agents because ugly UX is fine if the tool kills one daily annoyance. Curious which ones still survived after week two.

[Ash Explained @AshExplained]

@ali7ep @OpenAICodexCli @thsottiaux That smells more like state bleed than KV cache. I would check shared worktree paths or shared memory files first because cross-project collisions usually show up there.

[Ash Explained @AshExplained]

Reading Claude Code's settings docs pushed one boring rule higher in my stack:

Start with a deny list.

Block .env, secrets, credentials, build output.

"Be careful" inside a prompt is still a suggestion.

A deny rule makes the bad path impossible.

What files would you block before giving an agent your project?

[Ash Explained @AshExplained]

Reading Anthropic's hooks guide made one thing click for me:

If a step must happen every time, it should not live only in the prompt.

Prompts are for judgment.
Hooks are for rules.

Prompt: decide the fix.
Hook: run tests after edits, block protected files, re-add context after compaction.

If the agent is allowed to "remember later," it will forget on the run that matters.

What rule in your workflow should already be a hook?

[Ash Explained @AshExplained]

@Liz_LizLi @dexhorthy Yep. The why note matters more than the latest rule once the file gets long. I like a one line reason next to big changes. Do you keep that in the same doc or a tiny changelog?

[Ash Explained @AshExplained]

@andyglover Yeah, this works better when each model has a job. One pokes holes, one proposes moves. If both improvise, the good weird idea gets sanded off. Do you give them fixed roles or let them debate raw?