Drew Breunig
dbreunig.com · Personal blog · on HN since 2023
Drew Breunig writes about applied AI, including context engineering, prompts, DSPy, and how language models are used in practice.
What HN saysSimon Willison credits him with fantastic writing on context engineering, and commenters cite his posts on MCP, prompts, and LLM-as-judge.
43HN people shared it
11linked it in comments
143stories submitted
1.7Kstory points
3 months agolast shared
2023people per year2026
Most-shared posts
- 1How Long Contexts Fail9 people · 16 points · first shared 2025 · 5 comments on HN
- 2How to Fix Your Context5 people · 97 points · first shared 2025 · 26 comments on HN
- 3AI Use Cases: Gods, Interns, & Cogs4 people · 10 points · first shared 2024 · 1 comments on HN
- 4A OSS Library with No Code, Only Specs4 people · 2 points · first shared 2026 · discussion on HN
- 5Breaking Down Claude's 26k+ Token System Prompt3 people · 12 points · first shared 2025 · discussion on HN
- 6Using 'Slop Forensics' to Determine Model Ancestry3 people · 9 points · first shared 2025 · discussion on HN
- 7A Plea for Sober AI3 people · 8 points · first shared 2024 · discussion on HN
- 8Kimi K2 RL'ed Qualitative Data to Write Better3 people · 6 points · first shared 2025 · 1 comments on HN
- 9Let the Model Write the Prompt: Using DSPy to Decouple Your Task from the LLM3 people · 6 points · first shared 2025 · discussion on HN
- 10The Problem is Prompt Debt: You can't be model agnostic and hand-tune prompts3 people · 5 points · first shared 2026 · 2 comments on HN
- 11In Pursuit of Quiet AI Features and Better UX3 people · 4 points · first shared 2024 · 1 comments on HN
- 12Overfitting to First Party Harnesses3 people · 3 points · first shared 2026 · discussion on HN
- 13Cybersecurity Looks Like Proof of Work Now3 people · 3 points · first shared 2026 · discussion on HN
- 14Cat Facts Cause Context Confusion3 people · 2 points · first shared 2025 · 1 comments on HN
- 15Wikidata as a Giant Crosswalk File2 people · 153 points · first shared 2024 · 9 comments on HN
What HN people wrote when they linked it
thorum · 2026-01-09
My favorite is LLM-as-judge with a detailed rubric as discussed here: https://www.dbreunig.com/2025/07/31/how-kimi-rl-ed-qualitative-data-to…
dbreunig · 2025-08-07
Came here to say this: people present MCP’s verbosity as all the context the LLM needs. But almost always, this isn’t the case.
I wrote recently, “ Connecting your model to random MCPs and then giving it a task is like giving someone a drill and teaching them how it works, then asking them to fix your sink. Is the drill relevant in this scenario? If it’s not, why was it given to me? It’s a classic case of context confusion.”
https://www.dbreunig.com/2025/07/30/how-kimi-was-post-trained-for-tool…
simonw · 2025-06-30
I would classify AnyTool as a context engineering trick. It's using GPT-4 function calls (what we would call tool calls today) to find the best tools for the current job based on a 3-level hierarchy search.
Drew calls that one "Tool Loadout" https://www.dbreunig.com/2025/06/26/how-to-fix-your-context.html#tool-…
alexc05 · 2025-08-19
100% yes! There've been some other writers who've been doing parallel work around that in the last couple weeks.
https://www.dbreunig.com/2025/06/10/let-the-model-write-the-prompt.htm… is an example.
You can see the hands on results in this hugging face branch I was messing around in:
here is where I tell the LLM to generate prompts for me based on research so far
https://github.com/AlexChesser/transformers/blob/personal/vibe-enginee…
here is the prompts that produced:
https://github.com/AlexChesser/transformers/tree/personal/vibe-enginee…
and here is the result of those prompts:
https://github.com/AlexChesser/transformers/tree/personal/vibe-enginee… (also look at the diagram folders etc..)
foltik · 2026-04-07
Based on the recent leaks, their system prompt explicitly nudges the model not to do anything outside of what was asked. That could very well explain why it’s not fixing preexisting broken tests.
“Don't add features, refactor code, or make "improvements" beyond what was asked.”
https://www.dbreunig.com/2026/04/04/how-claude-code-builds-a-system-pr…
eichin · 2026-01-11
indeed - https://www.dbreunig.com/2026/01/08/a-software-library-with-no-code.ht… appears to be exactly that - the idea that the only leverage you have for fixing bugs is updating prompts (and, to be fair, test cases, which you should be doing for every bug anyway) is kind of upsetting as someone who thinks software can actually work :-)
(via simonw, didn't see it already on HN)
question_AK91 · 2025-11-20
https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-t…
This guy has a good write up on the topic
pants2 · 2025-09-17
Don't forget also that Cat Facts tank LLM benchmark performance: https://www.dbreunig.com/2025/07/05/cat-facts-cause-context-confusion.…
bobbiechen · 2025-09-03
(me again) and Drew Breunig just posted about the tensions of actually getting those ads in: https://www.dbreunig.com/2025/09/02/considering-ad-models-for-ai-produ…
tehlike · 2025-08-12
lukev · 2025-07-01
So, this is cool and a per-user memory is obviously relevant for effective LLM use. And major props for the temporal focus.
However, keeping a tight, constrained context turns out to actually be pretty important for correct LLM results (https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-t…
Do you have a take on how we reconcile the tension between these objectives? How to make sure the model has access to relevant info, while explicitly excluding irrelevant or confounding factors from the context?