a_c

My own set of heuristics

1. Use less subagents. The act of using subagent already needs dumping adequate amount of context for subagent to work effectively

2. Refactor when file grows large. Reasonably scoped files is easier to understand, for both human and agents, and consume less tokens because less guessing which section of the file to read

3. More capable models for planning, learning. Cheap models for execution. Same for human teams, senior member for higher level planning while junior member more on the act of translation of idea into code

4. I wasn't doing any fancy stuff that opus can't handle

5. Consult a second opinion for things you are not sure. (I have made a skill specifically for that)

6. Audit how your habit correlates with token consumption. I found out a skill a use frequently is using a lot of context just so a subagent can get all the context. (made a tool to audit claude specifically)

A natural extension would be a reverse steganography that trips AI into see other things in images

I’m also asking for my sons

Slightly off topic, anyone know of any good dinosaur illustration, ideally a large collection?

You are right! Completely missed the part mentioning specifically it was not claude code. Thanks!

Off topic, you could write "127.0.0.1 evil.com" to your /etc/hosts and bypass all the cloudflare thing I believe

Many such examples with variations

- browser extensions where a bookmarklet will do (media speed control)

- tolerate-able UX where alternatives available (e.g LLM chat interface without a table of contents is a pain to me, vibe coding an extension took me less than 1 session worth of credit)

- buying bulk vs buying individually

- doomscrolling vs reading a book

- standing vs tip toeing (one I discovered only very recently that I can tip toe any time to train my toes for climbing, no need dedicated training sessions)

- getting angry/depressed online vs realising most thing on internet is about provoking emotion

I climb and I want one :)

> talking about climbing apps with fellow climbers is a great way to be outside.

Indeed. The climbers I met are very supportive. They helped me scan the gym, shared their own climbing footage, which is what I'm trying to do, visualize climbing in a 3D scene.

I find it highly similar between running agents and running human teams.

Clear goal, share context, delegate but verify. Running a team of engineers also inevitably generates pages and pages of material, design spec, code, test, review. Just that we now do that with agents and agents are way less trust worthy

There are surprising portion of population expect a dedicated app to perform a particular function

You got me.

Indeed it is. I’m very grateful to what LLM enables me.

The revelation to me was that I used to code what I know, now I could code what I don’t know. The common path is that when I face something I don’t know, which is quite often, to move forward I have to level up my understanding.

Text to handwriting in 3..2..

I spent multiple 5-hour sessions spec-ing my climbing app with AI, clarifying interactions, algorithm, workflow etc. It ended up a frankenstein that I didn't recognise or know how each part interact with each other. Command line were a mess, different commands doing the same thing, with similar but redundant arguments. Everything looks kind of doing what I intended but overly convoluted and nothing really works. Real progress was made when I actually dig into the documentation of colmap/OpenMVS (essential tools, which I had never used before, in my workflow).

The AI gave me unprecedented turn around time in experimentation. The same experiments would easily take me over a month in the past. Now it was a few days. But still, real progress is made only when my understanding catch up with reality.

3D model construction of climbing gym. There is so much detail a normal climbing video can't show. Wall angle, hip position. I'm hoping to overlay human model, taken from climbing video, onto the 3D wall to make climbing visualization easier.

The system behaviour is totally up to anthropic's discretion. Its current behaviour is verifiable. In claude code, spawn a subagent with

1. Agent("Test")

2. look at your token usage

3. Repeat a few times

I didn't check again as I type this message but am somewhat sure subagent doesn't cache system prompt as of maybe last week

Every subagent send the same ~30k system prompts. If you are using fable/opus, that's easily 30% of a 5-hour window for 7 subagent, before doing any work

Last week I discovered apple's own photogrammetry library. https://developer.apple.com/documentation/realitykit/creatin...

I used it to reconstruct a 3D mesh of a bouldering wall. It worked perfectly. No gaussian splatting involved. I think the boundary between virtual and real world does get thinner by the day

Will give it a try thanks for the heads up. Mind if I ask if you are referring to SendMessage? I was testing on Claude code 2.1.196. SendMessage was not available. Skimming through their change log didn’t seem to have anything related to SendMessage. There is “ Fixed SendMessage silently misrouting when a re-spawned agent reuses a previous agent’s name — the tool now detects the mismatch and asks the caller to retarget” from 2.1.199. Not sure if they are related

Once you are popular, whatever said is automatically true

Regarding the lack of vision part, if you are using Claude or opencode, I've made a skill[1] that let's you talk with any models in Claude/opencode mid-session. You ask "Have claude opus to look at this PDF for a second opinion" during a session of claude with GLM5.2 or opencode with GLM5.2

It doesn't need to pass whole conversation history as context (unlike /model), you can ask follow up to that forked model (which sub agents in claude doesn't support AFAIK), and you can ask models from opencode while using claude.

[1] https://github.com/kmcheung12/second-opinion

I heard from someone before. charge more - the most vocal customers are the most demanding ones.

Also I wonder people tried charging for escalation (that's what enterprise plan is for I suppose)

Off-topic, I don't get why people still use a www subdomain, especially so in this case, www.map. Conway's Law in action?

Sometimes I find myself wanting the other way round, instead of letting LLM tell browser what to do, I want what I did manually in the browser be visible to LLM. I made a MCP to cover that use case

Traditional media is running their own LLM before LLM even existed. They absorbed their own bubble of information and hallucinate their narrative. I don’t mean it in a bad way. That’s how our tribal brain works, come up with stories with limited exposure to the world.

Must be the system prompts. Ask copilot to dump its system prompt, and compare the system prompt with claude. It is not accurate but handy. I bet they are quite different

While LLM models are bad at games, they are perfectly capable of writing a RL agent to train on the game itself.

Thinking human is rational is a highly irrational belief.

"Claude Sonnet 5 is available everywhere today at an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026. It then moves to standard pricing at $3 per million input tokens and $15 per million output tokens.2"

"Sonnet 5 is an upgrade to Sonnet 4.6, but it uses an updated tokenizer that changes how the model processes text to improve performance (this is similar to the tokenizer change we introduced with Claude Opus 4.7). The tradeoff is that the same input can map to more tokens: roughly 1.0–1.35× depending on the content type. The introductory pricing is set so that the transition to Sonnet 5 is roughly cost-neutral."

If we trust them, then it is roughly the same as sonnet 4.6

I haven't. Thanks for the heads up will give it a try! I use opus to comment on code design quite often though. It became a pattern that I made a skill for me to ask for second opinions https://news.ycombinator.com/item?id=48733092 Would love to hear your feedback if you don't mind!