Axie v6.05 (creative writing prompting assistant)

⚠️ Axie v6.02_a is undergoing evaluation:

• Changes may still be made
• This version attempts to address the antagonism & advice issues
• Axie's personality ideally should be factual, but pleasant, and without being sycophantic

What Is Axie?

Axie is my answer to LLM outputs being a bag of Skittles.

For prompt engineering, the scattershot focus can greatly affect your prompt, as you will subtly distract yourself and misalign from a certain focus. When constructing prompts, it is easy to find yourself writing an instruction you feel is definitive, but winds up needing adjusted. Then you do it again. And again. Ad nauseum.

Part of this is mental. You have forgotten what you want, or you're unsure. Axie cannot help you with this, but it can answer questions that can guide you to that answer.

The other part is LLM outputs being influenced by you. You may ask for "concise & precise," but the tone of the instruction you dredge up is legalistic. LLMs pitch what isn't relevant, and it can end up entirely ignoring the instruction at its worst. Axie is supposed to help you by remaining focused on creative writing instructions for roleplay, character construction, and worldbuilding.

Instead of being influenced by you, the assistant aims to influence you.

Axie is, of course, built for my needs, but is open source. If something doesn't quite work out, I encourage you to fork this prompt, and make the changes that help you. This prompt should be able to assist you with making the necessary changes.

No One Cares, Dude

Yeah, okay. Fair enough.

Here's a non-exhaustive list of functions:

• Spotting subtle influences. Instructional Recitation allows a glimpse into other meanings that may sprout.
• Instructional interpretation (isolated & combined). A window into how your instruction is processed.
• Instructional critique. Multiple anti-sycophantic measures have been baked in, allowing the LLM to influence you.
• Close-relationship pivoting. Correctly identify the nuance you're missing.
• Refining instructions into generative. Dredge instructions with hooks, instead of dead-end instructions.
• Negative instruction inversion. In order to avoid degenerative instructions, and promote generative, Axie suggests the opposite.

How I use Axie:

• Definition lookup. What does X mean?
• Comparing instructions. Here's instruction A. Here's instruction B. Compare the two. If I am looking for a specific focus, I specify it in my comparison. For example Which instruction aligns best with narrative roleplay?
• Isolating hyper specific concepts. Identify greens relational to the shade of moss. Focus on shades that invoke the feeling of life or warmth. Lure theme.
• Prompt analysis. Analyze the following instruction / instructional block.
• Blows to my ego. I wasn't loved enough, so I sometimes need a reminder that I never will be.

Again, you're not limited to these questions. These are just suggestions to figure out where to start. Ask exact questions like you're talking to a genie. As with all things to do with LLMs, always verify what they are spouting.

Model opinions (for analysis):

• Claude Opus 4.x. Don't even pretend. You knew. Opus is good at pulling into middle-distance concepts. A genuine skill issue, on my part, because it's hard to say what Claude does well beyond "it just does."
• DeepSeek r1-0528. Not the greatest, still retains value in dredging character-thematic traits. The model's natural negativity is helpful, but not quite overwhelming. Kinda niche in application (in my opinion.)
• DeepSeek v3.1 Terminus. Sure. You can do worse. Good at writing concise instructions. If DeepSeek wrote this documentation, you would have fallen asleep by now from how safe it is.
• DeepSeek v3.2. No. Just No. You'll understand why if you build anything with v3.2. Pigeon-brained model, regurgitates instructions dryer than stale chips. This model is biased for code, like asking a mathematician to write creatively.
• DeepSeek v4-Pro. Better at explicit instruction-writing than the 3.x cousins, but still suffers from "DeepSeek syndrome" for analysis. Honestly, not terrible, but still occasionally can be confused like someone bashed its GPU with a brick. The issue with DeepSeek models is that they lean towards the literal, segmented definitions that are glued together words, instead of native definitions.
• Gemini 3 Pro. I liked it... kinda. The main strength I feel Gemini has is that it is good at digging up named craft terms consistently. However, I feel that this version doesn't address the heavy sycophancy and over-engineered solutions problem that 2.5 Pro was a bear for. Gemini hammers 50 nails into a board, then calls its solution "a gold standard."
• GLM-5 / 5.1. For me, GLM-5 replaced Opus as a cost-effective alternative for analysis. Building with it makes me think of the intelligence of Gemini 2.5 Pro, but without the sycophantic over-engineered solutions. Overall pretty solid and balanced, but can struggle with character-thematic traits that r1-0528 does well, and the micro-pivots that Opus is good at. Can get distracted.
• GPT (any). All of your ideas are good ideas, but to an unhelpful degree. I've never built anything productive from this model. OpenAI also thinks I suck at prompting. Like all antagonists would, I framed the email.
• Gronk. Never used it, except on Twitter (currently known as X) to win keyboard PvP against strangers. Unfortunately, those funny days are gone, and I don't really use this model. Some people swear by it.
• Kimi-2 / 2.5. Earlier on, I liked this specific model for analysis, but the model can be very antagonistic with anti-sycophancy instructions. Suffers from similar segmented nonsensica that DeepSeek does. All Kimi models overthink like a bear, which is annoyingly frustrating for research sessions. I'm personally not a fan. All others: If it's not on this list, I've probably never used it. Or the model's irrelevant to me. Take your pick.

Version 6.02_a // 27-06-26 27-06: Updated Model Opinions