88Z AI, 22 November — @kloss_xyz asked Grok Bot to spawn two Cursor cloud agents for read-only audits with Gemini 4 Neon and Fable 5.2 at xhigh effort. The audit grades skills, flags merges or cuts already covered natively, and proposes fixes without applying them yet. Only edits agreed by both agents and Grok Bot are applied surgically, highest confidence first. The author reported consensus on 22 of 104 skills, 21%, and recommended rerunning when a new model drops. From the post, image and replies; not an 88Z test.
At a glance
- Have Grok Bot spawn two Cursor cloud agents to run read-only audits with Gemini 4 Neon and Fable 5.2 at xhigh effort
- The prompt grades every skill 0-10, flags merges or cuts, and asks for high-confidence fixes, but maps and plans before any edit is applied
- Apply only edits agreed on by both agents and Grok Bot, surgically, highest confidence first
- The author reported consensus on 22 of 104 skills, about 21%, and recommended rerunning after new model releases; that is the author's claim, not an 88Z test
Background: Why a skill library can lag a new model
In a post on 22 November 2026, klöss (@kloss_xyz) argued that the skills people attach to AI agents are often things the newest frontier models already do natively. Skill files do not automatically expire or shrink when a model improves. Old triggers, redundant steps, and outdated constraints can keep occupying context, and they may no longer be the right way to do the work.
The author calls this practice skill audits and treats it as recurring maintenance in a model's release cycle, not a one-off cleanup. The account below is based on the full post, the audit prompt shown in the attached image, and the author's replies in the same thread. The workflow, scoring rules, and the claim of 22 agreed edits are the author's own statements. They are not an 88Z reproduction or independent test.
Method: Have Grok Bot spawn two read-only cloud agents
The orchestration the author describes is specific: tell your Grok Bot to spawn two Cursor cloud agents, run them with Gemini 4 Neon and Fable 5.2, set both to xhigh effort, and have them perform read-only audits of the current skill library. The point of that pass is to find skills that the new model already covers natively, or that have fallen behind in their implementation. Read-only is the precondition, so the audit does not change the repository while it is still mapping problems.
In the screenshot attached to the post, Grok Bot replies that it is setting up the two read-only audit agents and confirms that Gemini 4 Neon and Fable 5.2 are both running at xhigh. That shows the opening of one run the author chose to publish. It does not establish that the same workflow has been verified on other accounts or setups.
What the audit prompt actually asks for
The author posted a prompt for the agents in the image and invited readers to reuse it. The prompt asks them to grade, rank, and score every skill in the current bot setup from 0 to 10. The criteria include triggers, correctness, learning from mistakes, bloat, safety, and future-proofing, plus other categories or subcategories the author may not already have names for, without overengineering the review.
It also asks the agents to decide which skills should be merged into others, or cut entirely because the model already does the work natively, and then to propose exact, highest-confidence fixes for all of them, not only the top scorers. The target, once the work is finalized, is a score of 9.5 or higher across skills, categories, and subcategories.
The most important constraint is that no edits should be executed in this pass. The prompt says to map and plan first, so the audit can be as accurate as possible and later surgical upgrades have the least chance of negative impact. What this stage should produce, then, is a complete scorecard, a set of keep-merge-or-cut decisions, and proposed fixes, not a landed diff.
Execution: Surgical edits based on tripartite consensus
After the audit, suggested changes are not applied automatically. The author describes a strict consensus rule: an edit is considered only if Gemini 4 Neon, Fable 5.2, and the orchestrating Grok Bot all agree on it. The application order is also explicit. Edits are applied surgically, highest confidence first, so the most certain fixes land before riskier ones.
The aim is to reduce the chance of introducing new errors. Agreement among the three systems does not prove that an edit will improve results on real tasks. It only means that, in this workflow, suggestions without consensus are left alone and are not written into skill files automatically.
The author's reported results, and what the thread does not settle
The author reported that all three systems agreed on edits to 22 of 104 skills, about 21% of the stack, and recommended running the audit again whenever a new model drops. Those figures come from the author. 88Z has not inspected the skill library and has not rerun the audit.
In the thread, one reader suggested using pstack and argued that it could replace a large share of existing skills. The author replied that pstack was already in use, and did not offer details when asked whether any optimizations had been found there. Another reader noted that agreement among three models shows only that the edits are plausible, not that they help, and asked whether any of the 22 changed skills had been rerun on real tasks to compare output. In the public replies retrieved for this article, the author did not give a substantive answer. The 21% consensus rate should not be read as a verified quality gain.
88Z Perspective
Viewed as engineering maintenance rather than a clever trick, the useful order is clear: keep the first pass read-only, score and plan before changing files, apply only high-confidence edits that more than one system agrees on, and repeat the audit when models change. For anyone who has already accumulated a large skill library, the risk is not necessarily a lack of skills. It is that old skills keep consuming context, or keep constraining a stronger model in outdated ways.
Human approval should still sit in front of deletions, mergers, and safety-related rules. Consensus is a filter against careless edits, not proof of better output. To decide whether the audit is worth turning into a routine, the missing comparison is how the same tasks look before and after the edits, which the original post does not provide. Readers who want to try the workflow should go back to the author's post and the prompt in the image, rather than treat this article as a licensed reprint or a hands-on test report.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- @kloss_xyzSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.