jcohen66

Input Caching

Not like regular output caching (think database).

Prompt Caching

- Cache only the input prompt not the output.
	- Model computes key value pairs at every transformer layer
	- Cache stores the computed kv pairs

Context Window Contents - Doc

• Summary

jcohen66

https://www.youtube.com/watch?v=xaQG-eD9d58&t=1s

Four Tactics To Stop Context Rot

1. Handoff Summary
2. Strategic files -> Text cheap, video expensive
3. Experiment -> Build intuitiion as models improve
4. Strategic Summaries -> Extend when you can't restart

Performance vs Context Usage

jcohen66

If you are using AI model as the Oracle you are in the wrong frame of mind.

1. Assign tasks don't ask questions. a) Describe the end state you want b) Provide the success criteria b) Let the system decide how to get there

2. Accept imperfections and start to iterate a) Ralph Wiggum works because it produces broken code b) Make it retry until it fixes it

jcohen66

Let the model write the code - Self acceleration loop - Use AI to make AI better

• Scope knowledge tasks AI is now preferred over humans

jcohen66

Unlike traditional software, they learn and adapt in realtime - Decide on actions after interpreting data - Not just following scripts - Attackers manipulate sensing, thinking or acting functions - They can hijack the whole process

Need to build in safeguards - Make sure it stays under my control - Continues to follow my instructions

jcohen66

From Daniel Miessler LI post

General Verifiability

The ability to verify or 'Hill Climb' on general tasks instead of just coding.

1. Verifiability is the key to progress.
2. Code is easy to verify, but we lack that for most other types of problems.
3. If we want to hill climb, we have to hill climb towards something.
4. That something is IDEAL STATE.

jcohen66

Deep Literature Review

Based on all the research papers I updloaded, create a comprehensive literature review covering: key arguments, methodological gaps, and underexplored research opportunities.

Cross-Document Connector

Compare the marketing strategy in Document A with user feedback in Document B, identify three potential misalignments between strategy and real user needs.

jcohen66

BRUTAL

B - Begin Fresh R - Right Model U - Use a Critic Persona T - Third Party Framing A - Ask Specific Questions L - Let It Grade Itself

Honest and Critical Feedback

The problem is AI wants to please you. It will return something like:

jcohen66

https://www.youtube.com/watch?v=EH8Mdx948Kw

Model

Reasoning mode enabled

• GPT-5.2
• Claude Opus 4.5
• Gemini 3

Ground

• ONLY documents

jcohen66

Get-WindowsCapability -Online | Where-Object Name -like 'OpenSSH*'