I don't believe you deserve the downvotes you're getting. Sometimes a good joke is so good that it appears shallow at first glance, but only upon reflection does the true humor show through. A shallow callout such as yours is sometimes necessary to call the reader back to it for that further reflection.
claude package has ten new versions published per week, and one new model every few months, one should definitely not rely on some undocumented tricks around it: it'll change, it'll break deep ultra-specific configurations
This is true, but also "temporal hacks" can make or break "cutting edge" workflows. I don't re-architect my claude instructions every release. But some releases justify examining your existing instructions and making sure they still fit the current model. And it has made a noticeable difference.
Probably. It does seem to have a built-in tool for exploring its own docs, and it has a special mode for working in a .claude/ directory. It's probably intended that users do this.
Is there an AI Coding Agent application structure emerging that is more or less universal across llm models? Is anyone collecting and writing on how to understand this architectural style?
The pattern across Claude Code, Codex and Cursor does seem to be converging: gather context, make a plan, execute, then verify.
What feels less standardized is how much control the user gets between those stages. Settings like showClearContextOnPlanAccept and disableAutoMode are interesting because they expose that boundary between “agent decides” and “human reviews before execution.”
That seems like the part where different coding agents will continue to feel very different in practice.
The blog post this discussion is for is one of the first in depth discussions I've seen of how these coding agents work. Most posts cover how to use them, not their internals and how they operate.
Yeah, I just had Claude fill out the task list, and then before hitting the end of the task list ask whether I wanted to continue or whether getting some of it done was enough....
Never. Ever. Ever. Tell Claude you have a deadline. It will do this on every task. It will half-ass things to “get it done in time” and argue about whether or not an approach will be done “on time” because it is estimating in human hours.
I decided to grep the actual binary to check. The current version is 2.1.156, but the post is based on 2.1.87.
Most of it holds up in 2.1.156: the hook response fields (updatedInput, permissionDecision, additionalContext, watchPaths, etc.), extra hook config fields (once, asyncRewake), skill/agent frontmatter (omitClaudeMd, criticalSystemReminder_EXPERIMENTAL, memory, color, context: fork), and autoMode/autoMemoryEnabled/autoDreamEnabled all show up as real Zod-schema config keys, not stray strings. autoMode has the allow/soft_deny/environment shape, plus an undocumented hard_deny.
Two things from the post I couldn't find in 2.1.156: yoloClassifier (the closest flag now is yoloEquivEnabled) and "Magic Docs" / the # MAGIC DOC: regex (the only MAGIC strings left are about file magic bytes).
I'm curious about that "magic doc" feature. Is that meant to go in CLAUDE.md or a project file? Does the file need to be mentioned during the session or does Claude automatically search for all mentions of the "magic doc" header in the project?
If software engineering were truly solved, like Anthropic claims, anyone could just vibecode it back. If only they stopped being allergic to the word "open" and open-sourced Claude Code, which, at this point, there is no practical reason not to.
There are numerous copies of the source available for claude code now that it has been leaked. The vast majority of what makes CC useful is already present, and it's unlikely that any killer features will be added going forward.
So it's already possible for someone to "vibecode it back". It's just perhaps not legal.
Clearly AI-generated writing (confirmed with Pangram). Amazed this has so many upvotes—are people even reading the article?
@dang I know you have so far resisted a rule for AI-generated content (as we now have for comments), but I personally would prefer a flag for articles so that I don't waste my time on slop.
surely concats of user input, stdout of external dependencies, and non-deterministic output feeding back directly to an eval is safe. it's never been a problem before. not even trying to check the boxes when it comes to security anymore.
Wow, not one mention of the env vars that have a far greater influence on how the models actually work under the hood - https://code.claude.com/docs/en/env-vars
Very important for bedrock deployments and other not-as-standard deployments
Key for how I've deployed it - disable adaptive thinking, max thinking tokens, disable telemetry, etc
Claude Code’s feature cardinality is breathtaking. At this rate, the next pope will be from Anthropic
given their recent speech, I’m sure theyre trying
sorry to post a shallow comment but this is a really excellent joke holy shit. well done.
I am sorry for my ignorance, but I don’t understand the joke. What does the pope have to do with this?
The Pope is selected by the College of Cardinals. The joke is a play on "cardinality."
also colah was invited by the literal pope to his encyclical https://www.anthropic.com/news/chris-olah-pope-leo-encyclica...
I don't believe you deserve the downvotes you're getting. Sometimes a good joke is so good that it appears shallow at first glance, but only upon reflection does the true humor show through. A shallow callout such as yours is sometimes necessary to call the reader back to it for that further reflection.
[delayed]
This article is 2 month old and partly outdated, Some features have been documented, for example the auto mode: https://code.claude.com/docs/en/auto-mode-config#define-trus...
claude package has ten new versions published per week, and one new model every few months, one should definitely not rely on some undocumented tricks around it: it'll change, it'll break deep ultra-specific configurations
in my experience, "undocumented tricks" break as often as documented features
like when they removed "clear context and execute plan" option after releasing 1M opus because "context window is not a problem anymore"
I so miss that clear context and execute plan mode! Now i have to keep clearing it manually again.
fyi you can re-enable it with `{ "showClearContextOnPlanAccept": true }` in ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/settings.json
also find `"disableAutoMode": "disable"` useful, since I'm typically switching between yolo and plan
It is possible to build automation that efficiently handles low level customization of new versions as they appear.
This is true, but also "temporal hacks" can make or break "cutting edge" workflows. I don't re-architect my claude instructions every release. But some releases justify examining your existing instructions and making sure they still fit the current model. And it has made a noticeable difference.
can i ask claude to generate its own config? "you're me, create your perfect set of config files."
Yeah I'd love to see a cookie cutter project with all of the best practices of the boilerplate files
Probably. It does seem to have a built-in tool for exploring its own docs, and it has a special mode for working in a .claude/ directory. It's probably intended that users do this.
There’s a slash command that’ll look through your conversation history to add allow permissions.
Is there an AI Coding Agent application structure emerging that is more or less universal across llm models? Is anyone collecting and writing on how to understand this architectural style?
The pattern across Claude Code, Codex and Cursor does seem to be converging: gather context, make a plan, execute, then verify.
What feels less standardized is how much control the user gets between those stages. Settings like showClearContextOnPlanAccept and disableAutoMode are interesting because they expose that boundary between “agent decides” and “human reviews before execution.”
That seems like the part where different coding agents will continue to feel very different in practice.
> Is anyone collecting and writing on how to understand this architectural style?
Are we on the same site? Is anyone writing about anything else?
The blog post this discussion is for is one of the first in depth discussions I've seen of how these coding agents work. Most posts cover how to use them, not their internals and how they operate.
> Honest status
> Not at 100% - and I want to be straight about why that's a longer road...
https://github.com/user-attachments/assets/961eff6c-0060-45d...
I just want Claude Code to stop giving up on achieving tasks. It's so annoying. Even with `/goal` or the new `ultracode` it gives up constantly.
My project is very complex (https://github.com/mohsen1/tsz) but Codex has no problem keep grinding without stopping like that
I just use /loop now which gives it a motivational prompt to keep the fuck going. Goal can be used too, but for some things a simple loop is better
Yeah, I just had Claude fill out the task list, and then before hitting the end of the task list ask whether I wanted to continue or whether getting some of it done was enough....
Never. Ever. Ever. Tell Claude you have a deadline. It will do this on every task. It will half-ass things to “get it done in time” and argue about whether or not an approach will be done “on time” because it is estimating in human hours.
Anyone else pick up on the fact that the article was released on Apr 01, 2026?
I decided to grep the actual binary to check. The current version is 2.1.156, but the post is based on 2.1.87.
Most of it holds up in 2.1.156: the hook response fields (updatedInput, permissionDecision, additionalContext, watchPaths, etc.), extra hook config fields (once, asyncRewake), skill/agent frontmatter (omitClaudeMd, criticalSystemReminder_EXPERIMENTAL, memory, color, context: fork), and autoMode/autoMemoryEnabled/autoDreamEnabled all show up as real Zod-schema config keys, not stray strings. autoMode has the allow/soft_deny/environment shape, plus an undocumented hard_deny.
Two things from the post I couldn't find in 2.1.156: yoloClassifier (the closest flag now is yoloEquivEnabled) and "Magic Docs" / the # MAGIC DOC: regex (the only MAGIC strings left are about file magic bytes).
I'm curious about that "magic doc" feature. Is that meant to go in CLAUDE.md or a project file? Does the file need to be mentioned during the session or does Claude automatically search for all mentions of the "magic doc" header in the project?
Have fun finding out that the undocumented feature you rely on suddenly stops working.
If software engineering were truly solved, like Anthropic claims, anyone could just vibecode it back. If only they stopped being allergic to the word "open" and open-sourced Claude Code, which, at this point, there is no practical reason not to.
There are numerous copies of the source available for claude code now that it has been leaked. The vast majority of what makes CC useful is already present, and it's unlikely that any killer features will be added going forward.
So it's already possible for someone to "vibecode it back". It's just perhaps not legal.
most of these are in fact documented, the rest either no longer exists, is still gated by feature flags (i checked), or has little use to most users.
you can however convince claude to create a local command with the extracted prompts for stuff like autodream
Clearly AI-generated writing (confirmed with Pangram). Amazed this has so many upvotes—are people even reading the article?
@dang I know you have so far resisted a rule for AI-generated content (as we now have for comments), but I personally would prefer a flag for articles so that I don't waste my time on slop.
What’s up with scrolling on that page?! I was locked into a page region several times until I finally gave up. Safari/Orion iOS current
Maybe programmed with Claude Code?
Substack is such a big platform that they should have resources to make sure their product works fine on common device/browser combinations.
I'm not sure how much leeway creators have in customizing their substacks to be honest (if they can use custom CSS for example).
I see the same with OpenAI's webpage on iOS.
Some of these things are documented. For example the model in frontmatter in skills is documented here: https://code.claude.com/docs/en/skills#frontmatter-reference
(It's not easy to find though, and lots of other docs doesn't mention it or link to this)
yup, it's super old
Related from then:
The Claude Code Source Leak: fake tools, frustration regexes, undercover mode
https://news.ycombinator.com/item?id=47586778
weak, even for you
That example classifier is horrendous. A simple substring search for ls/cat/echo/etc?
surely concats of user input, stdout of external dependencies, and non-deterministic output feeding back directly to an eval is safe. it's never been a problem before. not even trying to check the boxes when it comes to security anymore.
Can I do
And that'd be auto approved?still, far more effective than "NEVER FUCKING GUESS"
Wow, not one mention of the env vars that have a far greater influence on how the models actually work under the hood - https://code.claude.com/docs/en/env-vars
Very important for bedrock deployments and other not-as-standard deployments
Key for how I've deployed it - disable adaptive thinking, max thinking tokens, disable telemetry, etc
Article name: "everything the doc doesn't mention"
You: "they missed this feature that's in the docs !"
You're right that it's an important part of CC's config. But it doesn't fit the article's raison d'être.
Adaptive thinking isn't configurable if you're using 4.7 or higher though, so anybody on modern opus its basically useless.
https://code.claude.com/docs/en/model-config#adaptive-reason...
> Opus 4.7 and later always use adaptive reasoning. The fixed thinking budget mode and CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING do not apply to them.