AIvorax

Learning Al agents in public. Sharing what works, what breaks, and the prompts behind it

[noko @nokoxyz]

Anthropic just shared one of the best tricks for Claude Code.

The idea: run Sonnet, Haiku and Opus 5.5 together, each one doing only the job it's best at.

Claude Code has a hidden flag that puts Opus 5.5 on call while Sonnet does the work:

claude --model sonnet --advisor opus

it's Anthropic's "advisor strategy", built right into Claude Code.

here's the split:

> Sonnet 5.5 runs the session and writes the code
> Haiku 5.5 subagents read files and docs in parallel
> Opus 5.5 sits on call and reads the whole transcript when asked
> you stop paying Opus rates for routine turns

Opus usually steps in at three moments:

> before a plan locks
> when the same error shows up twice
> before calling the task done

Sonnet builds. Haiku explores. Opus reviews.

Anthropic's own numbers: Sonnet + Opus advisor scored 2.7 points higher on SWE-bench Multilingual than Sonnet alone, at 11.9% lower cost per task.

the full tree
> main session: Sonnet 5.5, edits code
> search: Haiku 5.5, maps the repo
> grep: Haiku 5.5, finds call sites
> docs: Haiku 5.5, pulls the API reference
> advisor: Opus 5.5, reads everything when consulted

paste this into Claude Code ↓

"Rebuild my Claude Code setup around this tree:
1. Make Sonnet my main model and set Opus as my advisor (advisorModel: opus).
2. Create three read-only subagents on Haiku 5.5: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
Show me every change as a diff first. No edits until I say go."

where it breaks, from my own runs:

> --advisor haiku on a Sonnet session gets refused. it keeps running with no advisor
> on a one-line question, the single Opus call cost 3x the whole Sonnet part
> it's still marked experimental and doesn't show up in claude --help

worth it for long refactors and bugs that keep coming back. skip it for quick questions.

[Bengi @Bengiyai]

Why do most people still use Claude Code like a chat window?

The ones getting real output run it as a small team. An agent reads the repo, another writes the plan, three edit in parallel, one runs the tests and one reviews. A human approves before anything merges.

Every agent works from the same spec in the same repo.

Investigate → Spec → Build → Verify.

Skills keep them from improvising: Superpowers for plans, TDD and systematic debugging, and Skill Creator for turning anything you repeat into a reusable skill.

Setup takes an evening. After that you give it a whole task and review what comes back.

Most people are still pasting one error at a time and waiting.

Full setup and copy-paste prompts in the reply 👇

[AIvorax @AIvorax1]

@Bengiyai Most people don't know how to use AI correctly, they use it as if it were a regular "chat bot" and that's the big problem

[AIvorax @AIvorax1]

@nokoxyz Usually, large prompts contain a lot of junk that prevents the AI ​​from understanding the task better, it is better to give 5 small prompts to the point

[AIvorax @AIvorax1]

5 AI tricks I wish I knew a year ago 👇

1.Skip the long prompt. Just say “Ask me 5 questions before you answer.” The result gets way better because the AI stops guessing.
2.Paste your draft and ask “What would a harsh critic say about this?” Much more useful than “improve this.”
3.Add “Explain it like I’m smart but new to this.” Kills the jargon and the baby talk at the same time.
4.Stuck? “Give me 3 different approaches and the downside of each.” You get options, not one random answer.
5.Finish any task with “What am I missing?” It catches the stuff you’d normally find out the hard way.

Save this and try one today.

[AIvorax @AIvorax1]

@xzolmoney I agree, this works better than just "make it better"

[AIvorax @AIvorax1]

@airadar_1 I think we still need people, if we don't limit AI it could end badly

[AIvorax @AIvorax1]

@JustinGorya I think they want to create a model that will be perfect for any task. Very soon we will know the truth

[AIvorax @AIvorax1]

@IndependentEco Try using this feature, it will be interesting to hear your opinion after some time

[AIvorax @AIvorax1]

If you keep hitting your Opus limits in Claude Code, you’re probably using it for way too much.

Try launching it like this:

claude --model sonnet --advisor opus

Sonnet does the actual work. Opus only gets called in when it matters: before a plan locks in, when the same error pops up twice, and before you call it done.

Want to push it further? Run your subagents on Haiku for the boring stuff like repo search, grep and diff checks:

CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model sonnet --advisor opus

Sonnet builds. Haiku searches. Opus reviews.

You keep Opus where it counts and stop paying for it everywhere else.

[AIvorax @AIvorax1]

@airadar_1 Your right 😅
But the bottom line is that work gets done better with less use of the limit

[AIvorax @AIvorax1]

@AI_cortexx I'm really looking forward to reviewing this MacBook to see how much these technologies will affect work and whether the price is justified

[AIvorax @AIvorax1]

@IndependentEco I think Anthropic will still win, but it will be interesting to compare these models right when they are released

[AInewssX @AInewssX]

Top-8 copy-trading looks genius for 30 days. Then this happens.

I simulated 20,000 traders. All of them were pure coin flips, with zero skill.

I ranked them by 30-day return and copied the top 8.

Past 30 days: $170 → $739. The top 8 showed +303% to +387%.

Next 30 days, same 8 traders: 6 of 8 lost money. My account ended at about $168.

Across 1,000 reruns, the median result was $168, and 54% of runs finished below the $170 start.

Why: with enough traders, someone always tops the 30-day chart by luck alone. Next month they're ordinary again.

Real copying is worse, because fees, a 0.3 second delay and skipped trades all cut returns.

Would you copy the top 8? Yes or no, and what would you check first?

Sources: Sergeenkov via The Defiant (84.1% of ~2.5M Polymarket wallets losing); Solidus Labs via CoinDesk (0.55% of profitable maker wallets took ~50% of gains).

[AIvorax @AIvorax1]

@xsmryze I wonder why no one thought of this before, with this solution the tasks will be performed much better

[xsmryze @xsmryze]

THIS IS AN ACTUAL F*CKING GOLDMINE

Anthropic just shared one of the best tricks for Claude Code.

The idea: run Sonnet, Haiku 5.5 and Opus 5.5 at once, each one doing only the job it's best at.

Better results and way fewer tokens, because every task lands on the right model.

You just launch Claude Code like this:

claude --model sonnet --advisor opus

here's the split:

→ Sonnet 5.5 runs the session, writes the code and applies the fixes
→ Haiku 5.5 swarms the repo in parallel: search, grep and diff audits
→ Opus 5.5 stays on call as advisor for the big calls and signs off on the diff
→ you stop burning Opus context on work that doesn't need it

Opus only steps in at three moments:

→ before a plan locks
→ when the same error shows up twice
→ before calling it done

Sonnet builds. Haiku swarms. Opus reviews.

JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split

the full tree
> Sonnet 5.5 runs the main session and coordinates the work
> search maps the repo on Haiku 5.5, read-only
> grep finds call sites on Haiku 5.5 in its own context
> audit checks the diff on Haiku 5.5 in a throwaway worktree
> Opus 5.5 on call as advisor, reviews the synthesized diff before sign-off
> JEV routes micro-forks in milliseconds

paste the tree and this prompt into Claude Code ↓

"Reconfigure my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three Haiku 5.5 subagents in .claude/agents: search (maps the repo), grep (finds call sites), audit (checks the diff).
3. Set isolation: worktree on audit so its checks never touch my working branch.
4. Add one rule to CLAUDE.md: delegate every repo search and test run to the Haiku 5.5 subagents to keep the main context small.
5. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go."

↳ https://t.co/YqkjQiqRGv

save it, it'll come in handy.

[AIvorax @AIvorax1]

@Mr_Salio What people once had to create by hand and take weeks to do, can now be done in a few hours with code, it's just amazing

[AIvorax @AIvorax1]

@UP_SHIFTED In English, it is easier to convey your thoughts through a small selection of words, while in Ukrainian, the same word in context can mean completely different things. Therefore, AI has to analyze the context better

[AIvorax @AIvorax1]

@NodePrint I agree with you, in my opinion AI is not good enough to analyze people and identify the perpetrator, but it can still track him down

[AIvorax @AIvorax1]

OpenAI just said something that sounds like a movie plot.

Someone hid a recording device inside one of their offices.

So they pointed their own AI models at it. What is it, where did it come from, who put it there.

According to OpenAI, it traced back to one employee who allegedly planted it only a few days earlier. The case went to the FBI, and the person was arrested and fired.

They say there’s no sign that model weights, source code or customer data were touched. The investigation is still going.

A couple of years ago we were arguing about whether AI could write a decent email. Now it’s helping catch someone bugging an AI lab.

Weird times. Would you trust AI to investigate something like this?

[AIvorax @AIvorax1]

@IndependentEco AI is a good tool as long as it obeys you