Your AI Gives Everyone the SAME Answer — This Free Skill Fixes It (ADHD for Claude Code & Codex)

[00:00:00]Open your AI coding agent right now. And ask it to design something. Anything. A rate limiter. A retry strategy.

[00:00:08]A name for that new service.

[00:00:10]I can already tell you what happens.

[00:00:13]It gives you the first idea it thought of.

[00:00:16]The same three or four textbook answers every other developer who asked that question got today.

[00:00:21]And here's the expensive part.

[00:00:24]That obvious answer feels right. So you build it. And a week later you find out it was the wrong one.

[00:00:30]Today I'm showing you a free open source skill that ends that.

[00:00:36]One command and your agent attacks your hardest problem from a handful of completely isolated directions at once.

[00:00:43]Hands you 30 plus genuinely different ideas. And this is the part nobody else does. Names which of those ideas are traps before you waste a week building one.

[00:00:55]First 30 seconds on why your AI keeps doing this.

[00:00:59]Because the fix only makes sense once you see the problem.

[00:01:03]Language models write one word at a time. And every word is steered by the words before it. So the first sentence the model commits to becomes an anchor.

[00:01:14]Everything after it mostly decorates that first idea. That's chain of thought. One line of thinking locked in early.

[00:01:22]Researchers tried tree of thought.

[00:01:25]Let the model branch into several paths.

[00:01:28]Better.

[00:01:29]But every branch still lives in the same conversation. The same context.

[00:01:35]Branch two can see branch one and it quietly falls in line.

[00:01:40]The anchoring survives.

[00:01:43]Which means this is not a prompting problem.

[00:01:46]You cannot ask the model to just be more creative.

[00:01:49]It's an architecture problem.

[00:01:51]And architecture problems need an architecture fix.

[00:01:55]That fix is this repo.

[00:01:57]It's called ADHD. Yes, really.

[00:02:01]A skill for coding agents by Udit Acharya.

[00:02:05]Open source MIT license.

[00:02:08]Built on the Claude and Codex agent SDKs and it crossed 1800 GitHub stars in about 8 weeks with a feature story on the New Stack.

[00:02:18]Here's the whole system.

[00:02:20]Two phases with a hard wall between them. Phase one, diverge.

[00:02:26]Instead of one model thinking in one context, it spawns five or more completely isolated agent calls.

[00:02:35]Each one gets your problem plus one deliberately weird point of view and each one is forbidden from judging its own ideas.

[00:02:44]The branches cannot see each other.

[00:02:47]Zero shared context.

[00:02:49]So, nothing anchors.

[00:02:52]Phase two, focus.

[00:02:54]A separate critic call running the opposite system prompt reads every idea cold.

[00:03:02]It scores each one on novelty, viability, and fit.

[00:03:08]It groups them by underlying approach.

[00:03:10]It flags the traps with reasons and it deepens the top three survivors into actual sketches with risks and first steps.

[00:03:20]Generator and critic are separate calls with opposite instructions.

[00:03:25]The separation is mechanical, not a polite request inside one prompt.

[00:03:30]That's the entire trick.

[00:03:32]Those weird points of view are called frames and they're my favorite part of this repo.

[00:03:37]A frame is not a persona. It's not pretend you're a marketing expert.

[00:03:42]It's a vantage that forces the model into a corner it would never drift toward on its own.

[00:03:48]15 ship built in. Listen to a few.

[00:03:51]The 3:00 a.m. on-call engineer.

[00:03:54]What design means you never get paged.

[00:03:57]The speedrunner.

[00:03:59]What glitches and skips crack this problem open?

[00:04:02]The competitor trying to break it.

[00:04:04]Attack your own design.

[00:04:07]The 10-year-old.

[00:04:08]Ignore convention.

[00:04:10]What's the naive approach? Inversion.

[00:04:14]Ask the opposite question.

[00:04:16]Then negate the answer.

[00:04:18]Remove the load-bearing assumption.

[00:04:21]What becomes possible if the database, the framework, the network just wasn't there?

[00:04:27]And one slot in every run is always reserved for a wild frame.

[00:04:32]Biology. Ant colonies. Markets.

[00:04:36]So, the output never collapses into sensible.

[00:04:39]Writing a frame of your own is about five lines of code.

[00:04:43]Here's what that buys you on a real problem.

[00:04:47]Straight from the repo's own benchmark.

[00:04:50]The problem.

[00:04:51]A command-line tool calls a language model.

[00:04:55]And it sometimes hangs for 90 seconds.

[00:04:58]Design the retry, timeout, and user experience strategy.

[00:05:02]The baseline single-shot answer gave four textbook patterns.

[00:05:06]Progressive timeouts.

[00:05:08]Exponential back-off.

[00:05:10]Hedged requests. Streaming with keep-alive.

[00:05:13]Then a sensible hybrid recommendation.

[00:05:16]Solid.

[00:05:17]It's the answer any senior engineer gives in 30 seconds.

[00:05:21]Which means it's the answer everyone already has.

[00:05:25]Now, the same problem through this skill.

[00:05:29]Over 30 ideas from six isolated frames.

[00:05:33]And the star pick was something the baseline never even considered.

[00:05:38]A rage-quit button.

[00:05:40]One click that instantly aborts the slow request.

[00:05:44]And resubmits it to a cheaper, faster model.

[00:05:49]Because maybe the real problem is that the slow model was the wrong model for that prompt.

[00:05:55]None of the four textbook patterns ever question the model itself.

[00:05:59]And on top of that, it flagged 20 seductive but broken ideas.

[00:06:03]Each with a one-line reason why it fails.

[00:06:07]That's the gap between competent and interesting.

[00:06:11]Quick pause.

[00:06:12]One frame. No tricks.

[00:06:15]If the last 2 minutes changed how you'll run your next design decision, that's what every video on this channel is built to do.

[00:06:23]Real talk.

[00:06:25]90% of you watching haven't subscribed.

[00:06:29]On Facebook, subscribe is the supporter button.

[00:06:32]That is what keeps the lights on here.

[00:06:35]On YouTube, it's free.

[00:06:37]And it tells the algorithm to send you more.

[00:06:39]Subscribe to support us.

[00:06:42]And if these breakdowns are saving you real hours, the join button on YouTube makes you a member of the lab.

[00:06:49]Okay.

[00:06:50]Now, the numbers.

[00:06:52]Because I don't ask you to trust vibes.

[00:06:54]The repo ships its own benchmark.

[00:06:57]Six open-ended engineering problems.

[00:07:00]This skill versus a plain one-shot answer from the same model, scored by an independent LLM judge, with a deliberately skeptical staff engineer prompt and the AB order randomized.

[00:07:16]The results.

[00:07:17]Breadth of ideas.

[00:07:19]9.0 versus 4.8.

[00:07:22]Nearly double.

[00:07:24]Novelty.

[00:07:25]7.8 versus 2.7.

[00:07:29]Almost three times.

[00:07:32]And trap detection.

[00:07:34]Naming the ideas that look great and fail in production.

[00:07:40]9.5 versus 1.8.

[00:07:43]Five times better.

[00:07:45]That's the biggest gap on the board and honestly the most valuable one.

[00:07:50]It won five of the six problems.

[00:07:53]Now the honest column because a repo's own benchmark is still the repo's own benchmark.

[00:07:59]On overall builder usefulness, it's basically a tie.

[00:08:04]7.7 versus 6.8.

[00:08:08]This tool does not make everything better.

[00:08:12]It makes ideation better.

[00:08:15]Even the New Stack's headline pointed out that outside experts want more proof.

[00:08:22]Fair.

[00:08:23]But one independent blogger already ran their own blind test.

[00:08:29]Two problems.

[00:08:30]Judge position swapped.

[00:08:32]The skill won both at a real cost.

[00:08:36]About 2.3 times the time and nearly twice the output tokens.

[00:08:40]So, no.

[00:08:42]This thinking isn't free.

[00:08:43]It's a few cents to a few dollars to widen one high-stakes decision.

[00:08:48]Which is cheap next to shipping the wrong obvious answer.

[00:08:52]Let's install it. You need exactly one command and it's on your screen right now.

[00:08:57]NPX skills add then the author's handle It auto-detects whatever agent you're running. Cloud code cursor codex client Gemini CLI Windsurf around 50 agents supported.

[00:09:15]From then on you can invoke it explicitly. /ad HD then your problem in quotes. Or just ask your agent for a few ways to do something and it auto-triggers on ideation questions.

[00:09:27]Prefer a standalone tool? There's a CLI npm install -g ad HD -agent and you can dial it. Three frames, eight ideas each.

[00:09:36]Keep the top two. And if you build your own agents, it's also a TypeScript library.

[00:09:41]One import, one run call, and you get back the short list. The non-obvious pick, the traps, and the clusters as structured data.

[00:09:52]One cost warning before you go wild.

[00:09:55]Every isolated branch reloads your base context.

[00:09:59]Your project instructions, your session state.

[00:10:02]Five branches, five reloads. So, treat it as a decision point tool, not an every keystroke tool.

[00:10:08]So, when do you actually reach for it?

[00:10:11]The repo gives a one-sentence test, and it's worth a screenshot.

[00:10:16]If a junior developer would Google the question and find the answer, the baseline wins. Don't spend the tokens.

[00:10:23]But, if a senior engineer would lean back and say, "Hmm, let me think about this differently for a minute."

[00:10:29]That is the moment this replaces.

[00:10:31]In practice, use it for architecture and design decisions.

[00:10:36]API and command line surface design.

[00:10:39]Fuzzy debugging, where you need hypothesis classes you haven't considered.

[00:10:44]Migration planning, naming things, and widening a code review beyond the checklist.

[00:10:50]Skip it for lookups.

[00:10:51]Bugs with a known root cause.

[00:10:54]Anything latency critical.

[00:10:56]And any question with exactly one correct answer.

[00:10:59]High stakes and open-ended, diverge.

[00:11:02]Everything else, just answer.

[00:11:05]One more thing made me take this repo seriously.

[00:11:09]Who's already shipping it?

[00:11:11]Repo wire, a mesh orchestrator, merged it into their default template.

[00:11:17]vendored it as their think plugin.

[00:11:21]WTF is my repo?

[00:11:23]Uses it as the explanation engine for onboarding people into unfamiliar code bases.

[00:11:30]It's even packaged as a plain text prompt in awesome prompts for people who can't install skills at all. And my favorite, an outside team ran a full evidence-based research review against the method.

[00:11:45]11 sources, eight validation rounds, and the author linked that critique from his own read me, and tracked the findings as public issues.

[00:11:55]A maintainer who advertises his own audit.

[00:11:59]That's exactly the energy you want in a tool you're about to trust with design decisions.

[00:12:06]So, here's the transformation plane.

[00:12:08]Yesterday, your agent gave you the answer everyone gets, and you hoped it was right.

[00:12:15]Tonight, one command later, it attacks every hard decision from directions you'd never think to try, argues with itself so you don't have to, and hands you the traps before you step in them.

[00:12:30]I've packed everything into one free PDF, the divergent ideation playbook, the exact install command for every agent, the when to use test card, all 15 frames, the five-line template for writing your own, and the honest cost math.

[00:12:50]Comment the word diverge and I'll send it to you free.

[00:12:53]Or grab it directly at hyperautomationlabs.co/free/diverge.

[00:12:58]Just getting started with AI coding tools?

[00:13:01]My step-by-step beginner guides are linked in the description.

[00:13:05]I post quick AI moves every day on Instagram and Facebook, and the full deep dives live here on YouTube.

[00:13:12]And if you want round two, me running this skill live on five real problems, receipts on screens, type run in the comments. Diverge. Go.