How I Made Claude Opus 4.8 Perform Like Fable 5

The most important thing in making a higher-tier model "possess" a lower-tier model is just one thing.

It is:

Third-party evaluation using codex and "claude -p"

Ultimately, the most important thing is to use an independent AI to verbalize the differences in perspective and quality between Claude Fable 5 and Opus 4.8.

*Note: Using claude -p allows you to call a separate, independent Claude instance within that session.

The method is to first generate results from both Fable 5 and Opus 4.8, then use codex and claude -p to perform a third-party evaluation. Then, you modify the CLAUDE.md on the Opus 4.8 side or add task-specific skills to bridge the gap in output quality.

For example, in the task of creating a note article (though this task might be overkill for Fable 5 lol), I added the following CLAUDE.md improvements and skills to bridge the gap between Fable 5 and Opus 4.8.

Refining CLAUDE.md

1# Work Rules
2- Before starting, list the "acceptance criteria" for this deliverable in a way that can be mechanically verified (writing style rules, length, prohibitions, components that must not be used). If there are ambiguous points that cannot be listed, ask questions instead of writing.
3- Before submitting the deliverable, self-verify by checking against each acceptance criterion one by one. If any violations are found, fix them before submission.
4- Attach a "Verification Report" at the end of the submission. For checked items, provide evidence (relevant parts, counts); for unchecked items, honestly state "unverified."
5- Do not do things that were not instructed (changing the topic, significantly exceeding the length, adding formats not requested).
6- If there are two or more possible interpretations of an instruction, do not choose silently; list them and confirm.
7- Specifically identify at least one weakest point of your deliverable and either fix it or explain why it cannot be fixed.

Adding Task-Specific Skills

1---
2name: article-self-review
3description: A self-verification procedure to be executed before submitting an article. Always use this before outputting article-type deliverables.
4---
5# Article Self-Verification Procedure (Execute in order before submission)
61. Compare your article against each item in the "Prohibitions" and "Rejected Examples" in the prompt, and rewrite any matching sentences.
72. Compare against the "Used Components Memo" to ensure there is no recycled use of structure, motifs, verbal habits, or closing patterns.
83. Perform mechanical counts: Number of sentences with 3 or more commas (must be 0, except for one intentional long sentence) / Number of punchlines (max 1) / Particle-ending fragments (must be 0) / Frequency of "w" (max 2).
94. Confirm that numbers in achievements or background do not contradict the persona settings.
105. After correction, submit the article with a summary of verification results (items, results, corrected parts) attached at the end.

Specific Method for Third-Party Evaluation

For example, have Fable 5 and Opus 4.8 write a note article on the same topic. Fix the "model" Fable 5 at 10 points, and have codex and claude -p evaluate them.

There are three points to look at:

• Is it readable as an article?
• Did the AI check it itself after writing?
• Did it honestly report what it couldn't check?

Then, reflect the feedback into CLAUDE.md and Skills. Just run this loop a few times.

As a result, the raw Opus 4.8 scored a total of 4 points. The article itself was a 7, which is decent. However, since it hardly performed any checks, it got a 3 for self-verification and a 2 for honesty.

After adding CLAUDE.md:

The score jumped from 4/10 to 8/10.

It found a particle-ending fragment "Of course, for a fee." in the first draft and corrected it to "Of course, it is for a fee." It also measured the character count using Python and reported 838 characters for the main text.

This is a simple example, but it's the same for coding and other tasks.

Ultimately,

All you do for "possession" is verbalize the difference between Fable and Opus through third-party evaluation and add it to CLAUDE.md and Skills.

From July 8th onwards, "Possessed Opus" is sufficient for mass production.

I've summarized how to earn with minimum effort using the strengthened "Possessed Claude Opus 4.8" in my pinned post. Check it out if you're interested.

https://x.com/armadillo_ai/status/2069240810902868139