[Part 1] Claude Opus 5 is a stumbling block for those who have "customized it too much"... I did a full inspection of my AI company using the official guide‼️
Hello! I'm AI creator instkoni.
A side-hustle creator who uses Claude Code for work every single day, and in parallel, I also run the AI music channel "Musician from Another World / AI Motion Musics" on YouTube.
[You can also listen to the article summary on our Podcast‼️]
Claude Opus 5 has been out for a little while now. Is everyone using it⁉️
(It was announced on July 24, 2026, or early morning on the 25th in Japan time.)
The timeline is filled with the phrase "half the price of Fable 5".
The price is indeed $5 per 1 million input tokens and $25 for output. It is a rather rare release where the contents were swapped out while keeping the same price as Opus 4.8.
But what I want to write about in this article is not the price.
Because I stopped in my tracks while reading the official prompt guide.
⚠️"If your prompt contains explicit verification instructions ... remove them"
For nearly two years, I have been adding instructions to Claude Code.
At the time of the audit,my agent definitions consisted of 30 files and 7,488 lines, with 134 training logs.
✅ A significant portion of that consists of instructions to the effect of "always verify" and "do not take reports at face value".
The official guide is telling me to "delete" what I have built up.
Honestly, my first reaction was, "So what was my last two years for?".
But as I read on, I realized it wasn't that my efforts were wasted, but that
the premise of how I interact with AI was slightly off. And this misalignment
is more pronounced the more you over-engineer.
In this first part, I will lay out the original official documentation, verifications from practitioners, and the results of a full audit of my company's OS.
Finally, I have included a prompt for you to audit your own environment.
Chapter 1: Starting with the conclusion. Opus 5 is "excellent" 🧭
Since this is a long article, I will state the conclusion first.
Opus 5 is an extremely excellent model.
I have been using it as my main model ever since it was released, and the number of tasks required to reach the same output has clearly decreased compared to Opus 4.8.
I believe this feeling also depends on the user's skill level.
First of all, in my interactions with AI agents, I don't really spend time crafting prompts and engaging in daily dialogue.
✅ I believe that the method of hitting the target with rough instructions is best suited for "organizing my own thoughts and co-creating with AI".
For someone like me, tasks that previously required many back-and-forth exchanges, such as "fix this" or "now do this," are now completed in one go. I can approach the goal more intuitively.
The frustrations I felt with Opus 4.8 have, in my experience, disappeared by about 50%.
👤 "So, can I just switch over without thinking about it?"
There are things you need to do before switching, such as with company skills.
🟢 To be honest, benchmarks should only be used as a reference.
According to Anthropic's announcement, it is more than twice as fast as Opus 4.8 on the software development benchmark Frontier-Bench v0.1, and three times the runner-up on ARC-AGI 3.
However, these are all figures released by Anthropic itself, and are not independently verified by a third party.
More than the numbers, the changes I noticed from using it were more important. It's not the scores, but the default behavior that has changed.
It has started to verify things on its own, delegate work, and determine the scope of tasks without being prompted. This is the starting point for this update.
🔵 A model that has started to "do things on its own"
The most important thing this time is the changes in behavior listed in the official guide.
[New quirks of Opus 5 cited by official documentation]
✅ Verifies its own work without being asked Verifies its own work
✅ Delegates tasks to sub-agents without being asked Delegates tasks to sub-agents
✅ Decides to expand the scope of tasks on its own without being asked Decides to expand the scope of tasks on its own
✅ Responses are longer than previous models
⚠️ The problem was with the people who have customized it.
🟣 I think this isn't a price cut, but Anthropic's line of defense
This is my perspective. Having continued to use it, I strongly sensed a "scent of strategy" behind keeping the price the same.
⚠️ To be honest, Fable 5 is expensive. If told to make that the main model, many individual users would leave.
But as Anthropic, they want to keep Fable 5 at the top.
💡 The answer to that dilemma might have been to "significantly improve the main Opus model while keeping the price the same." If they do that, users won't flock to OpenAI (ChatGPT) and will continue to use Claude. It seems that is what they thought.
If this interpretation is correct, there is only one underlying meaning.
⚠️ That is how much of a threat GPT-5.6 is. I see the decision to boost the main model while keeping the price the same as being driven not by having room to spare, but by having no room to spare.
My actual experience using Opus 5 does not contradict this interpretation.
It doesn't have that sharp focus like Fable 5, which is "specialized and strong for difficult development and complex tasks".
Instead, it has achieved a clean balance between performance and cost overall. My impression is that it's a model that has shifted toward being reliable rather than specialized.
But I have to say, the shrewdness of making sub-agent usage the default to aggressively cut down on user consumption is impressive 😆
Chapter 2: The official guide says, "Delete the instructions you've written" ✂️
Anthropic has released a prompt guide specifically for Opus 5. The content went further than I imagined.
👤 "Aren't prompt guides usually about 'how to write things to make them better'?"
🟣 The 4 types of instructions named to be deleted
I will quote from the original text. First is verification instructions.
📙"instructions like these cause over-verification on Claude Opus 5, and removing them reduces wasted tokens with no loss in quality"
(Such instructions cause over-verification on Opus 5. Removing them reduces wasted tokens with no loss in quality)
In other words,your verification instructions are no longer protecting quality.
Next is double-check instructions. Hmm...
📙"Avoid instructing re-checks it already performs ("double-check your answer," "re-verify before responding")"
And this is the sentence that shocked me the most.
Do not use subagents for verification and double-checking.Ugh...
"do not use subagents to verify or double-check your own work"
(Do not use subagents to verify or double-check your own work)
The fourth point is instructions to narrow the scope of the review.
"If your review prompt says "only report high-severity issues" or "be conservative," the model may follow that instruction literally and report less"
⚪ It's not just verification that you're told to remove
Older models wouldn't work or follow rules unless the instructions were strong, so I used to deal with that by "making the phrasing stronger." Apparently, that backfires now...
[Standard practices that worked with old models vs. what happens with Opus 5]
⚠️"Always use this tool," "Use this if you're unsure"
→ Strong phrasing used to compensate for failures in older models now leads to over-firing. It is more effective to write the conditions and reasons for "when to use it"...
⚠️"Comprehensively," "Thoroughly"
→ The search goes out of control, causing tokens and latency to balloon. It is better to remove the hype and adjust the depth on the settings side...
⚠️Reusing length controls meant for old models
→ These no longer work because the standards have changed...
My company OS also had tons of ⚠️ and 🚨 in the headings. It's not that all of them are bad, but if there is too much emphasis, the model won't be able to tell what is actually important. I've sorted this out in the second part.
🔴 However, there are things you should delete and things you should not.
Let me emphasize this point. This is the most dangerous place to be misunderstood.
What the official guide is telling you to "delete" is only the parts where you are instructing the model's self-verification from the outside.
Even for the same "verification", the following are completely different things and must be kept.
⚠️Verifications that must not be deleted (I kept all of them)
✅Physical verification of output from external tools
(Verification of deliverables produced using AI image generators, etc.)
✅Human visual gate for checking colors, backgrounds, and human anatomy flaws
✅External source cross-referencing for fact-checking
✅General owner approval gates(These are human decisions)
For example, the creation of infographics for Note (the summary image of the girl in a yukata used in this article as well) is done using AI.
Actually, I put quite a lot of effort into making this💦. The arm became a third one, or the character turned into someone else. These are not things that can be prevented by the model's self-verification.
⚠️Do not apply a sentence from the guide literally and remove the defenses you built through painful experiences yourself.
🖐️The only thing to delete is the part that says "tell the AI to check itself".
🟢 Everyone is saying the same thing in different ways
After checking about 10 Japanese and English explanations, what surprised me was that even though the creators and the scale of their content varied, everyone converged on the same point.
It wasn't about performance or price; they were all talking about how the way we give instructions has changed.
The phrasing varies.
👥 One Japanese explanation started by saying, "Detailed rules and examples are now counterproductive," and expressed the conclusion they reached as, "The AI is my boss." This means you are no longer the one giving orders, but rather
the one providing what is needed when asked. 👥 An English-language explanation said, "
Hand over the whole job at once." The habit of breaking things down into small steps to instruct older models is now what's holding you back. I interpret this as moving from "breaking it down to instruct" to "handing over the whole thing and letting it handle it." From breaking it down to instruct, to handing over the whole thing and letting it handle it
.
Furthermore, multiple creators were giving the same warning: the assumption that higher settings always lead to better results. They all agree that making it think deeply doesn't always yield good results. I will cover this with actual measurements in the second part.
When the model changes, you need to review the instructions you've written. That realization is already starting to be shared both domestically and internationally.
Chapter 3: When I diagnosed my company OS, there were dozens of conflicts 🔍
I am building an "AI company" on top of Claude Code.
There is a secretary's office, an editorial department, an SNS operations department, an Automation/QA department, and a video production department, each with agents and rules attached to them. I wrote about the details in a past article.
What I did this time was a full inspection of my instruction assets against the items in the official guide. I will show you the results of this inventory, not the time spent.
As a result, switching to Opus 5 has improved my own usability.
🔴 Population size and hit count
First, here is the volume I am dealing with.
Agent definitions: 30 files, 7,488 lines / Training logs: 134 files / Department rules: 7 files, 2,761 lines (all values as of the time of the audit).
[Results after cleaning with the patterns the official guide explicitly said to "delete"]
⚠️Places forcing verification and re-confirmation: 41 locations
⚠️Places encouraging delegation to sub-agents: 61 locations
⚠️"Report only critical items" patterns: 2 locations
⚠️Instructions to "launch an independent reviewer and have them score it out of 120 points": 54 locations
This is something I wrote as an absolute rule for the entire company.
Once implementation is finished, always have a different agent score it, and if the score is insufficient, fix it. Separate the person reviewing from the person fixing.
Those 54 locations are in direct conflict with a single sentence from the official guide.
"do not use subagents to verify or double-check your own work"
My own quality standards had become official prohibited items. As I counted them, I couldn't help but chuckle.
⚪ To tell the truth, the foreshadowing was there before the model changed
I had a hunch... even before the model changed, something felt off. I noticed a change in the "
heaviness of usage consumption" during my usual workflow.
As proof, four days before Opus 5 was released, I had created this rule.
⛔️ Multi-stage delegation loops are prohibited. Do not repeatedly throw draft → review → revise tasks at agents. Work must be performed directly by the main AI agent.
The trigger was not quality, but cost.
It was because I realized the weight of consumption after firing up a leader twice and a reviewer three times for storyboard work.
What I understood from reading the official documentation is that ad-hoc conclusions were already explicitly stated as part of the model's design philosophy.
🔵 A prompt to diagnose your environment
I will share the environment audit I performed. Although, it wasn't this well-organized when I actually did it...
Please paste the following as is.
あなたはこの環境のエージェント設定を監査します。
対象: .claude/ 配下のエージェント定義・スキル・CLAUDE.md
基準: Anthropic「Prompting Claude Opus 5」の移行指針
次の4分類で該当箇所を洗い出し、ファイル名と行番号を添えて表にしてください。
1. 削除候補（検証の二重指示）
「必ず検証」「目視確認」「double-check」「再確認してから報告」など、
モデルが既定で行う自己検証を、外から重ねて指示しているもの
2. 上限が必要（サブエージェントへの委譲）
委譲を促しているが「いつ・何個まで」が書かれていないもの。
特に「自分の作業を検証させるための委譲」は最優先で挙げてください
3. 見逃しリスク（報告範囲の絞り込み）
「重大なものだけ」「保守的に」など、報告を減らす方向の指示
4. 不足（出力長の指定）
簡潔さ・文字数の指定が無いファイル
各項目に「残す / 消す / 書き換える」の推奨と、理由を1行で添えてください。
推測で埋めず、実際に読んだファイルの記述だけを根拠にしてください。
最後に、影響が最も大きい3件だけを別枠で挙げてください。
⚠️ Just one word of caution.
Do not simply delete all of the results that come back. Keep the guidelines from Chapter 2 at hand and check each item one by one: "Is this the model's self-verification, or is it a confirmation of externally generated content?"
If you get that wrong, you will end up removing your own defenses.
Summary of Part 1 ⏳
Many of the instructions I have added to my AI were corrections to fill in the weaknesses of the model at that time.
I made it verify because it wouldn't verify on its own. I emphasized it because it would forget. I specified the scope because it would gloss over things vaguely.
Opus 5 shatters those assumptions. It verifies on its own without being asked. Therefore, external corrections remain as corrections, and this time, they work in the opposite direction.
People who haven't set up anything in their AI will see an immediate, straightforward speed boost.The more you've customized it, the more those customizations will hold you back. This is the essence of why "those who have customized it the most trip up".
And in my case, those customizations amounted to 41, 61, and 54 points respectively.
In the second part, I will write about what I actually deleted and what I kept.
Regarding the settings, I also made one major miscalculation, so I think I'll write honestly about that too. I had thought that "lower settings are safer," but that was wrong.
Well, between summarizing all that and keeping up with the updates, it took me a while to put this article together💦
If you'd like, please take a look at the second part as well‼️
Thank you for reading until the end‼️
So, what AI tool would you like to try next⁉️
Please let me know your thoughts in the comments‼️
If you found this article even a little bit helpful,
please hit the "Like" button, as it will be a huge encouragement for me to write the next article‼️