Kensei

context engineer

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

good read, and the logs back you up. both misses were the same question, in both reps: the session serializer. the model opened only the sessions note, named the right top class, and missed the tag class because the note never listed it. it never opened the source.

the rule I gave it was "answer from the note, open the source only if the note doesn't have it", and the note looked complete enough to trust.

your variant is the right next test: same notes as a map, source required before naming classes. my guess is it fixes that question and keeps a good part of the saving, since the note still points straight at the right file.

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

@0xGhostyx good guess, but no. /compact was the runner-up at -60%. the one that won doesn't summarize anything, it just drops the chat

[Kensei @0xKense1]

Same question, same right answer: 544,000 tokens the first time, 99,000 the second.

Between the two I typed one command, after a 30-turn chat with Claude Code. Nothing else changed.

You already have this command. It ships with Claude Code.

Which one is it?

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

@dhalishiva close, it came second. /compact cut 60% on the next 3 questions, but the summary call itself cost about 162k tokens after 30 turns. the winner cut more and cost nothing to run

[Kensei @0xKense1]

@jcwinter wow, that is great!

[Kensei @0xKense1]

Same question, same right answer: 544,000 tokens the first time, 99,000 the second.

Between the two I typed one command, after a 30-turn chat with Claude Code. Nothing else changed.

You already have this command. It ships with Claude Code.

Which one is it?

[Kensei @0xKense1]

@ajs6888 you are totally right

[Kensei @0xKense1]

Daisy Hollman, an engineer on the Claude Code team, at Code with Claude: "The fastest way to make your agent better at your code base isn't a smarter model, it's a tighter feedback loop."

Her idea is red squigglies for agents: a hook that nudges Claude at the moment of the mistake, and that Claude can override when it knows better.

· 00:00 - everything you customize is just text files
· 00:23 - what an IDE for Claude looks like
· 01:36 - red squigglies for agents
· 02:12 - PostToolUse hooks
· 02:38 - the generated-file example
· 03:18 - a tighter feedback loop
· 03:48 - tools that scale with intelligence

I used a hook the other way, as a hard block on network calls. It stopped all 33 runs that went for the network, and the agents almost never looked for another way around it.

Wall or nudge: which one are your hooks?

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

@Charlie01748449 think same

[Kensei @0xKense1]

@0xMorlex great to hear that

[Kensei @0xKense1]

@eth_ansh hilarious

[Kensei @0xKense1]

Daisy Hollman, an engineer on the Claude Code team, at Code with Claude: "The fastest way to make your agent better at your code base isn't a smarter model, it's a tighter feedback loop."

Her idea is red squigglies for agents: a hook that nudges Claude at the moment of the mistake, and that Claude can override when it knows better.

· 00:00 - everything you customize is just text files
· 00:23 - what an IDE for Claude looks like
· 01:36 - red squigglies for agents
· 02:12 - PostToolUse hooks
· 02:38 - the generated-file example
· 03:18 - a tighter feedback loop
· 03:48 - tools that scale with intelligence

I used a hook the other way, as a hard block on network calls. It stopped all 33 runs that went for the network, and the agents almost never looked for another way around it.

Wall or nudge: which one are your hooks?

[Kensei @0xKense1]

@talemi appreciate it, did that because I understand that a lot of people are not even aware of those tips, and this could help them

[Kensei @0xKense1]

Daisy Hollman, an engineer on the Claude Code team and once a chair on the C++ committee, spent 20 minutes at Code with Claude on where your Claude Code tokens actually go.

Her rule for the context window: "Don't pay for what you don't use."

· 00:00 - the context window is a box
· 02:03 - npm on an Arduino
· 03:43 - the KV cache: change something early, pay 10x for the rest
· 06:42 - MCP: 20 servers x 15 tools fill the prompt
· 11:50 - skills: the description is always loaded
· 14:20 - hooks: zero token cost until they fire
· 17:02 - subagents
· 18:38 - why plugins can't ship a CLAUDE.md

Her CLAUDE.md point matches my test: 16 KB of rules in CLAUDE.md cost 46% more tokens per question, and the same rules in a skill cost 31% less. In my setup a bare "ok" already costs 19,100 tokens.

How much does your setup load before you type anything?

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

@daniel_mac8 Thx mate, appreciate it, did really hard research

[Kensei @0xKense1]

@AlexWithAi_ same lol. what eased mine: in my tests about half the runs with no rule tried to go online on their own, and a single hook blocked 33 of 33. so now auto mode only runs with that hook on. you on auto or full yolo?

[Kensei @0xKense1]

The Claude Code team explains in 7 minutes how Anthropic runs its own agents safely

Simon Willison: "I still mostly run Claude Code in YOLO mode and feel incredibly guilty about it."

· 00:11 - the YOLO mode confession
· 00:33 - almost everyone at Anthropic uses auto mode
· 00:45 - thousands of evals and hired red teamers
· 01:27 - "It doesn't catch 100% of things"
· 01:50 - a Sonnet classifier checks every bash call
· 02:42 - when a network request tries to leave the sandbox
· 04:07 - why you shouldn't build your own AI Slackbot
· 05:05 - giving Claude its own credentials
· 06:07 - credentials the agent can use but never see

The interesting part: even the team that built auto mode says it doesn't catch everything.

So I tested my own setup: 600 Claude Code runs and the 4 layers that stopped every attempt.

[Kensei @0xKense1]

https://t.co/JGPa3emUtl

[Kensei @0xKense1]

The built-in general-purpose one, Claude picked it on its own in 10 of 10 runs. The work copy had no CLAUDE.md, so the only setup each worker repeated was the skills list (30 skills). Delegating cost about 45k tokens on the task call and saved about 54k on the follow-ups, so it came out near zero. A leaner worker could tip that, worth its own run. How much did it save you per call?

[Kensei @0xKense1]

Hitting your Claude Code limit is mostly a skill issue.

I measured 12 popular tips on 945 runs. 5 of them saved nothing: subagents, the caveman skill, deny rules, turning off MCP servers, effort low on Sonnet.

The 7 that held up fit on one page. Skip #1 and you pay almost 5x for the same answer:
1. /clear between unrelated tasks: -79%
2. /compact on the same task: -60%
3. long rules in a skill, not CLAUDE.md: -31%
4. effort low on Opus: -17%
5. a flat notes map: -32%, but 18/20 right
6. naming the file: 20/20 right vs 8/20
7. Haiku for lookups: about 12x cheaper

Score yourself: how many of the 7 do you already do?

[Kensei @0xKense1]

https://t.co/WUk7rWCYuZ

[Kensei @0xKense1]

@0xMorlex try it best

[Kensei @0xKense1]

Hitting your Claude Code limit is mostly a skill issue.

I measured 12 popular tips on 945 runs. 5 of them saved nothing: subagents, the caveman skill, deny rules, turning off MCP servers, effort low on Sonnet.

The 7 that held up fit on one page. Skip #1 and you pay almost 5x for the same answer:
1. /clear between unrelated tasks: -79%
2. /compact on the same task: -60%
3. long rules in a skill, not CLAUDE.md: -31%
4. effort low on Opus: -17%
5. a flat notes map: -32%, but 18/20 right
6. naming the file: 20/20 right vs 8/20
7. Haiku for lookups: about 12x cheaper

Score yourself: how many of the 7 do you already do?