Vampa

@polymarket and crypto researcher. AI contributor. editor at @doubletop_io

[Vampa @0xVampas]

20 minutes that explain harness, loop engineering, evals and llm ops end to end. no code, no math.

the part everyone skips: a loop is not engineering until you say where it stops.

"it needs to know when it should stop"

the harness is the same idea one level up. the model is the horse, everything you built around it is the reins.

all five nodes are on one short:
Prompt → Context → Harness → Loop → Graph

book it, watch it twice, then paste it into your Claude and watch what changes ↓

[Vampa @0xVampas]

New Boris Cherny’s Opus 5 playbook is 10 page long.

it opens with him deleting 80% of Claude Code’s system prompt and shipping it anyway:

“the model is actually a little bit more intelligent without these prompts”

he gave one prompt 15 days ago and hasn’t touched it since. it’s still running right now.

the stack:

Prompt → Context → Harness → Loop → Graph

step 1 → delete the entire system prompt on release day. not trim. delete
step 2 → run the real product on real work. never guess what the model needs
step 3 → restore a line only after the same failure shows up twice
step 4 → price every line you keep. the model reads it on every single call
step 5 → your CLAUDE.md, skills and hooks die every 6 months too

the receipts:
→ Bun’s runtime rewritten Zig to Rust in 11 days. one prompt, ~100k lines, in production today
→ nobody asked it to, but the desktop rewrite opened its own Slack channel to live-blog its progress
→ 30 loops now maintain 4 codebases daily. one of them is literally called abstraction police
→ CLAUDE_CODE_SIMPLE=1 strips every prompt including tool prompts. undocumented. go try it

book now, read tonight. tomorrow you delete 80% of your own prompt ↓

[Vampa @0xVampas]

@0xChaseTM Thanks dude

[Vampa @0xVampas]

loop engineering is prompt engineering with the human deleted.

17 minutes, free. everything you need on harness and loops, and nothing you already know.

02:52 → the 12 hour task that lies about finishing
04:06 → what harness actually is, in three parts
12:59 → a loop inside a loop inside a loop
16:36 → all 6 components on one screen

watch now, then read how to run this on opus 5 ↓

[Vampa @0xVampas]

New Boris Cherny’s Opus 5 playbook is 10 page long.

it opens with him deleting 80% of Claude Code’s system prompt and shipping it anyway:

“the model is actually a little bit more intelligent without these prompts”

he gave one prompt 15 days ago and hasn’t touched it since. it’s still running right now.

the stack:

Prompt → Context → Harness → Loop → Graph

step 1 → delete the entire system prompt on release day. not trim. delete
step 2 → run the real product on real work. never guess what the model needs
step 3 → restore a line only after the same failure shows up twice
step 4 → price every line you keep. the model reads it on every single call
step 5 → your CLAUDE.md, skills and hooks die every 6 months too

the receipts:
→ Bun’s runtime rewritten Zig to Rust in 11 days. one prompt, ~100k lines, in production today
→ nobody asked it to, but the desktop rewrite opened its own Slack channel to live-blog its progress
→ 30 loops now maintain 4 codebases daily. one of them is literally called abstraction police
→ CLAUDE_CODE_SIMPLE=1 strips every prompt including tool prompts. undocumented. go try it

book now, read tonight. tomorrow you delete 80% of your own prompt ↓

[Vampa @0xVampas]

@0xChaseTM very usefull course

[Vampa @0xVampas]

@SkadooshGG yep bro

[Vampa @0xVampas]

loop engineering is prompt engineering with the human deleted.

17 minutes, free. everything you need on harness and loops, and nothing you already know.

02:52 → the 12 hour task that lies about finishing
04:06 → what harness actually is, in three parts
12:59 → a loop inside a loop inside a loop
16:36 → all 6 components on one screen

watch now, then read how to run this on opus 5 ↓

[Vampa @0xVampas]

@0xMortyx More people really scary about ai replace humans. And her words getting calm a bit

[Vampa @0xVampas]

@0x1mpulse great article bro, booked

[Vampa @0xVampas]

@de1lymoon RAG is a revolutionary tool for AI

[Vampa @0xVampas]

@SkadooshGG thanks dude, its really gamechanger for your work with opus 5

[Vampa @0xVampas]

New Boris Cherny’s Opus 5 playbook is 10 page long.

it opens with him deleting 80% of Claude Code’s system prompt and shipping it anyway:

“the model is actually a little bit more intelligent without these prompts”

he gave one prompt 15 days ago and hasn’t touched it since. it’s still running right now.

the stack:

Prompt → Context → Harness → Loop → Graph

step 1 → delete the entire system prompt on release day. not trim. delete
step 2 → run the real product on real work. never guess what the model needs
step 3 → restore a line only after the same failure shows up twice
step 4 → price every line you keep. the model reads it on every single call
step 5 → your CLAUDE.md, skills and hooks die every 6 months too

the receipts:
→ Bun’s runtime rewritten Zig to Rust in 11 days. one prompt, ~100k lines, in production today
→ nobody asked it to, but the desktop rewrite opened its own Slack channel to live-blog its progress
→ 30 loops now maintain 4 codebases daily. one of them is literally called abstraction police
→ CLAUDE_CODE_SIMPLE=1 strips every prompt including tool prompts. undocumented. go try it

book now, read tonight. tomorrow you delete 80% of your own prompt ↓

[Vampa @0xVampas]

Everyone is posting Opus 5 benchmarks.
Boris Cherny posted something else.

"the success rate for prompt injection attacks drops to ~0"

That's with defenses layered: alignment, injection probes,
Auto Mode in Claude Code.

Put it next to the other thing about Opus 5 and it checks its own
work and keeps going until the task lands.

A model that verifies itself and can't be talked out of the task
is the first one you can actually walk away from.

[Vampa @0xVampas]

New Boris Cherny’s Opus 5 playbook is 10 page long.

it opens with him deleting 80% of Claude Code’s system prompt and shipping it anyway:

“the model is actually a little bit more intelligent without these prompts”

he gave one prompt 15 days ago and hasn’t touched it since. it’s still running right now.

the stack:

Prompt → Context → Harness → Loop → Graph

step 1 → delete the entire system prompt on release day. not trim. delete
step 2 → run the real product on real work. never guess what the model needs
step 3 → restore a line only after the same failure shows up twice
step 4 → price every line you keep. the model reads it on every single call
step 5 → your CLAUDE.md, skills and hooks die every 6 months too

the receipts:
→ Bun’s runtime rewritten Zig to Rust in 11 days. one prompt, ~100k lines, in production today
→ nobody asked it to, but the desktop rewrite opened its own Slack channel to live-blog its progress
→ 30 loops now maintain 4 codebases daily. one of them is literally called abstraction police
→ CLAUDE_CODE_SIMPLE=1 strips every prompt including tool prompts. undocumented. go try it

book now, read tonight. tomorrow you delete 80% of your own prompt ↓

[Vampa @0xVampas]

@0xMortyx it`s really usefull skills. i`ll use them

[Vampa @0xVampas]

IBM just dropped a new 45-minute course on agent loops and why yours still isn't shipping faster:

0:58 – chatbots know. agents do
11:56 – they felt 20% faster. they shipped 20% slower
17:17 – the fix vibe-coding can't do
21:31 – "almost right, but not quite"
39:27 – the agent found what nobody wrote down

45 minutes of this beats most paid agent courses.

Watch it immediately and save to bookmarks

[Vampa @0xVampas]

IBM just dropped a new 45-minute course on agent loops and why yours still isn't shipping faster:

0:58 – chatbots know. agents do
11:56 – they felt 20% faster. they shipped 20% slower
17:17 – the fix vibe-coding can't do
21:31 – "almost right, but not quite"
39:27 – the agent found what nobody wrote down

45 minutes of this beats most paid agent courses.

Watch it immediately and save to bookmarks

[Vampa @0xVampas]

Everyone is posting Opus 5 benchmarks.
Boris Cherny posted something else.

"the success rate for prompt injection attacks drops to ~0"

That's with defenses layered: alignment, injection probes,
Auto Mode in Claude Code.

Put it next to the other thing about Opus 5 and it checks its own
work and keeps going until the task lands.

A model that verifies itself and can't be talked out of the task
is the first one you can actually walk away from.

[Vampa @0xVampas]

Everyone is posting Opus 5 benchmarks.
Boris Cherny posted something else.

"the success rate for prompt injection attacks drops to ~0"

That's with defenses layered: alignment, injection probes,
Auto Mode in Claude Code.

Put it next to the other thing about Opus 5 and it checks its own
work and keeps going until the task lands.

A model that verifies itself and can't be talked out of the task
is the first one you can actually walk away from.

[Vampa @0xVampas]

@Serantych This race for benchmarks is already boring everyone

[Vampa @0xVampas]

Everyone is posting Opus 5 benchmarks.
Boris Cherny posted something else.

"the success rate for prompt injection attacks drops to ~0"

That's with defenses layered: alignment, injection probes,
Auto Mode in Claude Code.

Put it next to the other thing about Opus 5 and it checks its own
work and keeps going until the task lands.

A model that verifies itself and can't be talked out of the task
is the first one you can actually walk away from.

[Boris Cherny @bcherny]

Opus 5 is a great model for coding, data analysis, design, biology, knowledge work.

More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.

And when layering defenses -- strong model alignment, combined with prompt injection probes, combined with Auto Mode in Claude Code -- the success rate for prompt injection attacks drops to ~0. This is new and exciting! More about this soon.

https://t.co/Tc7z2FqJhQ

[Vampa @0xVampas]

@SkadooshGG Great article dude!

[Alex @de1lymoon]

https://t.co/4aVb5Mr21l