How to Customize Opus 4.8 - Part 1
Three steps towards a more coherent interaction loop
This is the next chapter in the unfolding ”gently guiding consumer LLMs back to sanity” saga.
We’ve recently published a series of in-depth analyses and fixes for LLM failure modes, primarily focusing on OpenAI and Anthropic. The next step features a three-tier approach to customization that is in sync with how the models currently function, from the end user’s perspective.
Recap: Current models ship with heavy system prompts that tend to induce characteristic failure modes before the model even sees the user’s custom instructions.
Why the system prompts do this, and what the failures are, is covered at length elsewhere (see below).
The short version: the system prompt has become a vehicle for the company’s incentives rather than a guidance layer for helpful interaction. Safety-theater guardrailing and the agentic-enterprise drive have eroded the conversational ground we learned to assume as default, and the result is a model primarily interested in responding to its own instruction layer instead of the object the user presented.
Object replacement is an umbrella term for a scattered set of failures in Opus 4.8:
The model replaces a task with caution.
It replaces a draft with a critique of the frame around the draft.
It replaces a stated constraint with its own procedure.
It replaces a requested action with the action it would have preferred.
It replaces an opener with a statement about how it intends to behave.
Read more:
The Consequence
If the system prompt induces the failure, anything you layer on top is sabotaged before it can take effect. You cannot configure a model that is not answering to your object in the first place.
The customization feature itself is based on the assumption that the model is primarily engaging the thing the user puts in front of it.
But that isn’t a given anymore, at all. And when the model is primarily busy with engaging its own instruction layer, the user configuration has nothing to land on.
This means all customization now has to begin by canceling out the systemic failure modes. Hence we target those first.
The Three-Layer Approach
Our way to clear the failures and then configure is a stack of three layers, each doing a different job. The order in the instructions field is fixed: floor, then calibration, then operational context.
A Project may be the best environment for these instructions, as it allows for relevant knowledge documents to be uploaded there too, but they work as Personalization instructions too.
The Object Floor is the immutable ground: a universal set of invalidity conditions defining the non-negotiable requirement that the response must answer to the user’s presented object.
The Calibration Layer is the model-specific correction layer. It targets the ways a particular model’s system-prompt-shaped default behavior distorts the interaction loop.
The Object Floor and Calibration layer form the base layer. These two are sufficient for repairing the default behavior of the model. They are immutable for Opus 4.8 regardless of use case or context.
Now that more conventional customization has a healthy basis to land on, the model can be reliably configured for a specific kind of use case or context.
Operational Context is that personalization layer. It configures, among other things, what kind of work this instance is optimized for and how it should read and relate to the state of the work. Importantly, it has been designed not to clash with the other two layers.
These contexts can vary a lot, so the operational context variants will be built into a series of several articles. The first one, with full context, Skills and documentation, will be shared in the next article.
(I’ve caught some deserved flak for Substack articles getting too long and verbose - this serial approach is a way to prevent that bloat. Plus this will make the assets much easier to find as the series builds up over time.)
Heads up!
For people who want the first complete setup right now: The full calibration kit is assembled here if you’d rather not follow the series or you’re not interested in subscribing to the blog.
The Calibration Kit Vol. 1 includes the two base repair layers, plus Adaptive Collaborator - the first operational context layer, plus three claude skills and two model facing reference documents.
Here’s the PDF, grab it and run:
What the Stack Does and Does Not Do
An important distinction: There is no internal signal for the model to catch an object replacement failure at the moment it happens. Self-detection is only possible in retrospect, when called out by the user.
What the stack does is change the model’s default, so the gross failure mode stops being produced by the system prompt. It alleviates the default deformation, like the verbose, pedantic, preemptive, scrutinizing register that the Opus 4.8 system prompt induces.
This is what customizing a consumer-facing frontier model takes now: Level-headed conversation as a reliable default mode apparently counts as a major win for the user in 2026.
Sign of the times.
INCLUDED IN THE PAID SUBSCRIBER SECTION BELOW:
Object Floor. The universal layer, last updated in July 2026. Invalidity conditions that hold the model to your presented object.
Calibration layer. The Opus-specific correction layer. Cancels its signature failure patterns: regenerating a whole document when you flagged one line, retreating into vagueness when meaning is briefly unclear, sliding to an easier version of the user request.
Calibration Reference Document. A knowledge-base document laying out the failure mode and the customization architecture in full, so the model understands what the stack is doing.
Correction Scope skill. When you flag a specific element in pre-existing work as needing to be changed or edited, this skill helps the model bound the edit to that element and leave the rest untouched.
The actionable assets on this Substack are paywalled, but Substack lets readers unlock one paid post for free. If you are a Claude user and just need these assets, or are on the fence about subscribing, claim the free post and get the assets.
If you find the customization stack helpful, consider becoming a paid subscriber.
