If you have a CLAUDE.md file, it is probably working against you right now.
I can say that with confidence because last week Anthropic deleted more than 80% of the instructions that make Claude Code work.
Nothing broke. They realized they were limiting Claude.
Think about your team’s harness and CLAUDE.md file (or AGENTS.md file). A smart engineer wrote each of the lines there, on purpose, to fix a real failure in a model that genuinely needed the help.
They were right at the time.
Then the model got better. And the instruction stayed as junk context.
Source: Anthropic, The new rules of context engineering.
I ran /doctor on my own setup this week to see how bad mine was. It read my last 50 coding sessions and came back with a duplicate install, a permission mode I never updated, and a stack of rules I wrote in a different era for a different model.
Six months of accumulated care, working against me.
Deleting it took twenty minutes, and now I’m blazing through coding sessions with Opus 5.
In this week’s newsletter, we get into the 6 new rules of context engineering, with 1 prompt you can use to start optimizing your Claude Coding today with the new Generation 5 models, and my takes on Dario’s distillation stance.
What’s Inside This Week:
- ALIGN: Kimi K3 costs 94 cents a task, Jensen Huang’s first tweet ever, and a swarm that rebuilt SQLite for $1,339
- BUILD: All six new context rules in plain English, what to do about each one, and the only thing I found that survives the next release
- CULTURE: Who owns the language, and why Anthropic’s distillation argument does not add up
ALIGN: The Week Everything Got Cheaper
The most interesting things I found this week in AI.
Listen first: Why the Best AI Teams Stopped Chasing Models
Weekly AI Clarity, episode 44. 67 minutes. 90% of companies cannot point to a single sustained dollar of value from AI. The ones that can are seeing a 3.7x median return.
Jonathan and I break down the Kimi K3 release and the US versus China model race, then get into the AI gateway pattern that Stripe, Uber and DoorDash use to control token spend, the “company brain” from YC’s latest RFS, and how you actually get a whole org to adopt AI.
Open the episode → or subscribe on Spotify and Apple.
Anthropic Deleted 80% of Claude Code’s System Prompt and Nothing Broke
Six new rules for context engineering on Claude 5 models. Give judgment instead of rules. Use progressive disclosure instead of front-loading. Stop compounding in CLAUDE.md.
Most of the advice circulating about prompting was reverse-engineered from models that no longer exist. If you built your team’s practices in 2025, you are running a 2025 harness against a 2026 model and paying for the privilege in tokens.
A Chinese Model Now Costs 94 Cents Where Claude Costs $1.80
Moonshot AI shipped Kimi K3 on July 16. A 2.8 trillion parameter open weights model that debuted at number three on the Artificial Analysis index and took the top spot on the Frontend Code Arena. Demand got heavy enough that Moonshot paused new subscriptions four days later.
The benchmark position is the least interesting part.
The interesting part is that for the first time, an engineering leader building a token budget has a credible answer for the middle of the distribution that is not a frontier SKU.
Every comparison post you will find on this is measuring the wrong thing. Nobody is actually asking which model is better.
Here’s the most important question to answer: what percentage of your team’s tasks genuinely need frontier intelligence? Ten? Forty?
Have you ever measured it?
Cursor Rebuilt SQLite Six Different Ways and the Bill Ranged from $1,339 to Five Figures
Same task, same 835-page manual, same held-out test suite, 100% pass. Frontier model as planner plus a cheap model as worker came in around 8x under frontier-only. In the winning run, the planner produced a small share of the tokens and about two thirds of the cost.
Very cool stuff. As a builder I can appreciate their thorough testing and rigor! The takeaway to me is that the expensive model does not need to do the rote execution work. It needs to collapse the ambiguity with planning, and then cheap models just follow instructions.
Boom. That is most of your savings, and it costs nothing to try!
Jensen Huang’s First Post Ever Was a Lobbying Letter
“Open Weights and American AI Leadership.” Signed by Microsoft, Meta, Hugging Face, Palantir, and 20 others. Within a day the list doubled to about 50, picking up OpenAI and Google.
Three pages, and one of them is logos.
It still does something Sam and Dario have not managed in two years of testimony: it explains why this matters to you as a person, to the company that would hire you, and to the country, in that order. We have had a thought leadership vacuum at this altitude.
A GPU chip salesman who used to work at Dennys filled it. I will take it. I love me some Jensen Huang. We’ll dive into this more in the Culture section.
Claude Code Now Writes Its Own Harness
Dynamic workflows let Claude write a short JavaScript program on the fly to spawn and coordinate subagents, each with an isolated context window. Jarred Sumner used it to port Bun from Zig to Rust.
Oh man. Jonathan and I have been hand-rolling this exact thing for months and now it ships in the box.
Six named patterns come with it: classify and act, fan out and synthesize, adversarial verification, generate and filter, tournament, loop until done.
BUILD: The Six New Rules, In Plain English
Anthropic published six rules. Each one is written as a swap: stop doing the old thing, start doing the new thing.
Source: Anthropic, The new rules of context engineering.
Below is each rule, what it actually means, why it works now when it did not before, and the specific thing to do about it.
Start here, though.
Do this first: run /doctor
Before you form an opinion about your own setup, run /doctor in Claude Code.
It reads your install, your plugins, your MCP servers, your CLAUDE.md, and your recent sessions, then tells you where you are wasting context. Mine went through 50 sessions to make its recommendations.
Then paste the context engineering article into the same session and ask whether your setup is ready for the Claude 5 generation. My first pass without the article was thin. With it, the report got specific about which of my rules were dead weight and why.
Do this before you cut anything by hand. Your opinion about your own config is six months old. Mine was.
One caution before you start: if other engineers work in this repo, tell them first. Their commands and skills are about to change underneath them.
Continue reading
Get the full newsletter, free.
Join founders and builders who read Self Aligned every week.
Rule 1: Give judgment, not rules
What it means: Replace prescriptive commands with the principle behind them.
Anthropic’s own example. They used to write “never write multi-paragraph docstrings or multi-line comment blocks.” Now they write “write code that reads like the surrounding code: match its comment density, naming, and idiom.”
Why it works: A rule is a guess about every future situation, made in advance, by someone who cannot see the situation. The model can see the situation. When your rule and its judgment disagree, your rule wins and the output gets worse.
Your move: Search your CLAUDE.md for the word “never.” Each one you find, ask what you were protecting against, then write that instead.
Rule 2: Design interfaces, not examples
What it means: Stop pasting sample outputs to establish a pattern. Put the intent into the shape of the tool itself.
A status parameter that accepts pending, in_progress, and completed tells the model how to use it. Three examples of you using it tells the model less, and costs more.
Why it works: Examples narrow the search space. The model treats your three samples as the boundary of what is allowed instead of the center of what is wanted.
Your move: Find the example blocks in your skills. Delete them. Make the parameter names and enums carry the meaning instead.
Source: Anthropic, The new rules of context engineering.
Rule 3: Disclose progressively, do not front-load
What it means: Stop putting everything in the opening context. Load detail at the moment it becomes relevant.
In practice: move detailed guidance out of CLAUDE.md and into Skills that get invoked when needed. Structure long skills as a tree of files rather than one wall of text. Tools now load on demand through search rather than sitting in context from the first token.
Why it works: Everything you load up front competes for attention with everything else you loaded up front, including the actual task.
Your move: Take your longest skill file. Split it. Leave a short index at the top and push the detail into files that get pulled in when they apply.
Source: Anthropic, The new rules of context engineering.
Rule 4: Say it once, in the tool description
What it means: Guidance about a tool belongs in that tool’s description. Nowhere else.
Most of us wrote it twice: once in the tool description and again in the system prompt, because the model used to need the reminder.
Why it works: It does not need the reminder. The second copy is not reinforcement. It is a second voice that can drift out of sync with the first.
Your move: Grep your system prompt for tool names. Every hit is a candidate for deletion.
Rule 5: Let auto memory do it
What it means: Stop hand-curating CLAUDE.md as your project’s long-term memory. Claude now persists relevant context on its own.
Why it works: You were guessing at what would matter later. Auto memory decides based on what actually came up.
Your move: This is the one that stings. A lot of us spent a year treating that file as institutional memory. Strip it back to what the article recommends: what the repo is for, the gotchas, the non-obvious constraints. Delete anything Claude can work out by reading the file tree.
Rule 6: Use rich references, not simple specs
What it means: Prose descriptions are the weakest way to tell Claude what you want. Point at real artifacts instead. Test suites. An implementation from another codebase. HTML mockups. And rubrics.
Why it works: A description of a thing is lossy. The thing is not.
Your move: Next time you would write a paragraph describing the desired behavior, link the test that proves it instead.
The prompt that does all six for you
Reading six rules and hand-auditing your own repo are different jobs. Paste this into a fresh Claude Code session and it will do the second one.
First, make sure I am on the latest Claude Code.
Check `claude --version` in a terminal, not from inside this session.
A running session will happily tell you it is current when it is not.
Then run /doctor, and read this in full:
https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
Now audit my harness against those six rules. Read every CLAUDE.md in this
repo, every file in .claude/skills/ and .claude/commands/, my MCP server
list, and my recent session history.
For each rule, give me:
1. The specific lines that violate it, quoted, with file paths
2. Why that line costs me now instead of helping
3. The exact replacement text, or DELETE if it should just go
Rank the list by tokens saved per session.
Two things I care about more than a clean report:
- If a line is load-bearing for a real failure I would hit again, say so
and leave it alone. Do not delete guardrails that are still earning
their place.
- If cutting something changes behavior in a way I would notice, flag it
before you touch it.
Do not edit any files yet. Give me the plan first.Two notes on that prompt. The version check is real: on the pod this week Claude told me I was on 2.1.220 while the footer said 2.1.217, so verify from a terminal. And the do not edit yet line matters more than it looks, because the first pass will want to delete about half your CLAUDE.md and you want to read that list before it happens.
What should you actually invest in, if the models keep absorbing your work?
Now the harder question. If I just deleted six months of careful work, what do I build next that does not evaporate the same way?
Jonathan’s answer, on the pod, unedited:
“Focus on the rubrics. You know, focus on the domain specific rubrics… Everything else, the orchestration level layer, is just going to get like acid washed away, by these models as they get better and better.”
Acid washed away.
Everything else may be moot by the next model release.
But those grading rubrics that you carefully version and curate… The ones that represent your particular compounded domain expertise and intelligence…
Those compound, and are durable.
Let me make that concrete, because grading rubric is a word that sounds like process theater until you see it.
Think about your high school English teacher.
She has 45 essays to grade and one weekend.
Essay quality is subjective.
She knows good writing when she reads it, and she cannot say why in a way that survives a parent phone call.
So she publishes a rubric before anyone writes a word. Argument construction, so many points.
Grammar errors under a threshold. Bibliography present and formatted. Structure.
The rubric does two things at once.
It makes her taste legible to someone who does not have it.
And it makes her grading defensible to someone who was not in the room.
That is the whole game right now.
Claude Code and Codex compete on exactly this: teams of people annotating outputs, building evals, evolving rubrics for what counts as a good pull request.
Anthropic is not shipping better vibes every single release.
They are shipping a better definition of done, refined against millions of sessions.
What you have that they do not, is the domain niche tacit knowledge and expertise.
You know what a good output looks like in your niche, on your codebase, for your customers, in a way that no general harness will ever encode. That knowledge lives in the heads of your best people and leaks out the door when they leave.
So write it down as a rubric. Version it in the repo.
Treat it as a first-class asset with the same care you give schema migrations.
When your agent finishes a task, it should be able to grade itself against a standard you wrote, and you should be able to look at that grade and trust it.
Build In Public
$10M ARR
I see a line of sight to an enterprise case study.
And that is a stepping stone to PMF.
What tells me that?
We’re getting signal on the problems we’ve determined are real problems, and our solution hypotheses to solve for them for our customers and prospects.
I see it in my meetings, and I see it in my content.
I also believe, as of right now (knock on wood), we have sufficient runway to get there without signing additional customers.
I am hopeful we hit our first ARR end of year then keep rolling next year. Just gotta keep grinding.
Climb V10, Run 100 miles
I’ve been dragging ass the past week and haven’t been climbing or running as much, much to my own chagrin. Lots of work going on, and I’ve been dealing with some emotional stress on the family side which feels heavy.
Anyone else have issues setting and enforcing boundaries with family?
It’s tough, but I have all the tools and the fortitude. Just gotta keep improving at my skill in doing this, regulating, and making progress on all of my goals for the business and myself.
Lately, I have not been trying as hard at climbing as I normally do. I gotta admit my climbing stoke has been low. It’s odd because I feel quite stoked about everything else.
I find my inner critic has been loud, telling me I should be doing more. I feel bad about it.
And I then constantly remind myself, it’s okay to not be trying hard at touching rocks and plastic for a bit.
Much of the time, local minimas and maximas feel like everything—but when you zoom out, it’s really nothing.
So, carry on!
Life is still dope.
In terms of my running stoke…
I have a trip coming up, where I’ll be working from my van for a few weeks going to the Eastern Sierras in California, then up to Oregon, Washington, and ultimately Squamish in Canada to meet friends for a bouldering trip.
I’m really excited!!
Kenji and I have been itching to get to the mountains again. It’s been about a month since our last excursion. Can’t wait to go to questing with my boy. (:
Dog Dadding
Speaking of Kenji, lately I’ve been taking him on BART and testing his training there. He gets so nervous and anxious and drools a bit because of it. But he stays under my seat and me, and is a good boy overall.
What a good boy. (‘:
CULTURE: Who Owns the Language
Today’s Culture section gets a bit more philosophical.
What is right and wrong?
We humans fight wars and argue until our faces are red and our throats are sore, about what is right and wrong.
And the AI era is no different.
There’s an argument that Anthropic is fortifying its position around against distillation, the process that Chinese labs use to produce models such as DeepSeek, GLM, Kimi, etc.
Dario, CEO of Anthropic, argues for cracking down on industrial-scale distillation.
And yet I see irony in this. Why?
Two reasons.
One, the frontier labs took the corpus of human knowledge and human language, trained on it, and compensated nobody.
Now open weight models distill those results back out and hand them to everyone.
Two, Anthropic themselves can require KYC to verify identity and raise the bar for China’s industrial scale distillation. They have the means to do this, but are not.
Why not?
They have an impending IPO, and growth matters most for that. They’re not going to block people from paying them. So they’ll keep parroting the narrative and playing the regulatory capture game by poking at the government to do something about it.
That’s my take anyway.
Jonathan thinks of open weights as closer to electricity than to a nuclear weapon.
Jensen Huang seems to agree on the need for open models in the world, per his recent letter.
We’ll have to see how it plays out.
Keep building, Robert
P.S. Run **/doctor** and hit reply with what it found. I want to know whether six months of drag is typical or whether I was uniquely sloppy. I read every one.
P.P.S. If you are sick of the noise, subscribe to Weekly AI Clarity.
Every week Jonathan and I give you the AI news that actually matters with real takes, the memes that got us through the week, and the tactical meta we are testing on ourselves. News plus laughs plus what to actually do on Monday, in one sitting.