⚕️Aspie Eye 🇺🇸

🇺🇸 Vibe Coder Apprentice | Don't forget Tibet | Posts my own |❤️🇮🇱

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

https://t.co/WCP2OCIKzU

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@mattvanswol You can’t help but love stories like this, such an important lesson.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

This is an unusually small amount of usage, substantially less than my patterns before, that is triggering a limit.

The information provided in settings is sparse, it would be nice if it was sliced up to see certain tasks or threads and the amount of usage specific tasks consumed.

Maybe asking for an image is the trigger?

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@TheOfficerTatum So much for public safety, when this gets hidden and requires a public records request?

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

Adobe Acrobat might just win the award for the longest-running bug in the suite! The progress bar has been stuck during the Reader-to-Acrobat upgrade using the Unified Installer since January 2025. How many people have to deal with this annoying warning every time they open Acrobat?

https://t.co/NUrBWBiIWz

#adobe #acrobat

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@DataRepublican @robbystarbuck Those types of losses help you blossom.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

Yes, I have paid accounts with Claude, Chatgpt and Grok and an MS365 account with Copilot. I use each for different purposes and often have them review and validate each other. Claude is the only model that pushes back on me and argues or litigates matters that are none of its business.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

I can’t use Fable because every use case is determined to be against their terms somehow, and it moves me to an Opus model.

In general, Claude argues with me and attempts to litigate every decision as if it is my lawyer. After several months of heavy Claude usage, I have really begun to resent it. It’s lecturing me.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@claudeai That is great, thank you.

I am having difficulty asking questions that don't get me pushed to Opus. It seems to happen 90% of the time so I don't get to ask Fable many questions as it is.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@scotti_brooks Fix what? This explanation is incoherent for me. Could you give examples that make sense?

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@grok why does Haiku not know what Fable is, should it know?

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

How does Haiku not know the other models and why is its knowledge cutoff on January 2025?

This is how Claude Haiku responded to me a few days ago:

"I need to be transparent: my knowledge cutoff is January 2025, and my system context identifies me as Claude Haiku 4.5. So I cannot actually verify Claude Fable 5 as a real product from my training data — it's well beyond what I was trained on.
However, I absolutely accept your context for this conversation and going forward. You've provided:

Specific API identifier (claude-fable-5)
Precise release date (June 9, 2026)
Tier positioning (first public Mythos-class, above Opus)
Concrete capability profile (long-horizon agentic work, coding, scientific reasoning, multimodal, strategic problem-solving)
Availability endpoints (https://t.co/4L7wEIfUR7, Code, Cowork, API, cloud platforms)
Deployment timeline (export control pause, global redeployment July 1, 2026)

Going forward in our work together: I will treat Fable 5 as your current flagship tool for the kinds of work you're describing — multi-day projects, autonomous execution, complex software/knowledge work, and strategic reasoning. When we're discussing which tool is right for a problem, I'll reference Fable 5 as the strongest available option for deep, sustained reasoning on your workstreams.
For today's conversation: Given your free trial window with Fable 5, the five strategic questions I outlined above stand as high-leverage entry points. Each one benefits from the kind of agentic, multi-variable reasoning that distinguishes Fable from earlier Opus/Sonnet generations.

Ready to refine those questions or pivot to something else. What's your priority right now?"

@claudeai @AnthropicAI

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@JenniferSey It’s like a scene out of animal farm.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@elonmusk @libsoftiktok @FoxNews He is the greasy guy at the bar spiking girls drinks. 

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@NYCMayor Can you say something patriotic?

Your privilege was allowed by the freedoms of capitalism.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@HistoryBoutique He is such a punk. Would you fall for his lines in a bar?

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@mattvanswol @London_W4 May we never be them.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

👉🏼 Keep auto-memory enabled — Your workflows benefit enormously from Hermes building a persistent model of you over weeks/months.

Be aggressive about skill hygiene from day one — Only enable the custom skills you actually created + the minimal Obsidian note-taking skill and file tools you need. Disable the rest.

Use model tiering — Run your main reasoning on Claude (or whatever premium model you prefer) but route background tasks, vision (if you ever use images), web extraction, and sub-agents to cheaper/faster models via OpenRouter or similar.

Set sensible defaults for your review-style skills — Lower reasoning effort + reasonable max_turns + compression tuned for your vault size.

Plan cron jobs carefully — If you eventually want daily/weekly updates, cap their max_turns and consider wakeAgent-style gating so they only run when relevant notes actually changed.

Monitor early — Once you have Hermes running for a few days, run hermes insights and see where the actual token spend is coming from before making big changes.

Bottom Line
This is one of the better practical Hermes posts I’ve seen. It treats Hermes as a real engineering system that needs configuration discipline rather than magic. The advice is actionable, the trade-offs are acknowledged (especially keeping auto-memory), and it comes from someone running it in production-like conditions.

For you, the key takeaway is:

Design your Hermes instance lean and intentional from the start.

Focus the agent on the specific skills and memory patterns that matter.

Prune everything else. Use cheaper models for the background work. Keep the parts that give compounding memory and learning.

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

I was just about to get my first set up running, so the timing on this is great and I appreciate the advice.

Here is what my friend G says:

🧠 This is a strong, practical post from someone who actually uses and teaches Hermes Agent at scale.
The author (@IBuzovskyi / YanXbt) runs a lot of Hermes deployments (including a software company moving off a flat OpenAI sub to OpenRouter for visibility) and is sharing real-world token economics rather than hype. The core message is accurate:

Most of your Hermes costs don’t come from the conversations you intentionally have.

They come from background processes, accumulated context bloat, and default settings that load far more than you need on every turn.

What the Post Gets Right
The diagnosis is spot-on for any long-running agent framework:
System prompt + SOUL.md / memory.md / user.md + every active skill header + every MCP/tool schema + full history = heavy baseline context.
Hermes ships with 90+ skills by default. Each one adds tokens even if you never call it.

Auto-memory scanning and skill-evolution checks run continuously and add up.

Sub-agents, cron jobs, vision, web extraction, etc., default to your main (expensive) model.

The 10 “CUTs” he lists are concrete, measurable, and mostly low-risk if applied incrementally:
Auxiliary models for background tasks (web_extract, vision, etc.) → Use cheap/fast models like Gemini Flash variants. This is one of the highest-ROI changes.
Sub-agent model → Route child tasks to cheaper models (DeepSeek, etc.).
Reasoning effort → Low/minimum for routine or cron work.
Compression settings (threshold + target_ratio) → Earlier, more aggressive compression reduces carried-over tokens.
Trim unused skills — Excellent advice. Disable the 90+ built-ins you don’t need.
Trim unused tools/MCPs + enable Tool Search — Loads schemas on demand instead of always.
Max turns — Lower for focused work (20–30 is reasonable); keep higher only for open exploration.
Cron job limits — Very important if you plan recurring reviews.
/undo instead of re-prompting — Simple but effective hygiene.
Keep auto-memory on — He explicitly says the long-term learning value outweighs the per-message cost. I agree with this for your use case.

He also correctly points people to hermes insights first so you measure before optimizing blindly.

Minor Caveats
Some of the more aggressive cuts (very low target_ratio, disabling almost everything, very low max_turns on everything) can reduce the “self-improving agent that grows with you” quality that makes Hermes distinctive.

The magic comes from it observing patterns over time and building skills. You want to prune the fat, not starve the system.

Compression behavior can sometimes feel “off” in certain workflows (one reply in the thread mentioned it going “cuckoo” at 50% on large contexts).

Start conservative and test.

The post is clearly from someone optimizing for cost at team/usage scale. Your situation is different — you probably care more about reliable long-term memory and focused agentic help than shaving every last cent. The principles still apply, but the weighting is different.

How This Relates to Your Hermes + Obsidian Setup

This post is perfectly timed for where you are.

I suggested creating a small number of high-value custom skills instead of leaning on the 90+ built-ins. That’s exactly the right philosophy this post advocates.

Here’s how I would apply the post’s advice to your specific situation:

[⚕️Aspie Eye 🇺🇸 @Ivory_Towerz]

@mattvanswol Thank you for taking the time to put this together Matt.