Skip to content

Commit 5c8b42d

Browse files
committed
Update README to clarify prompt caching benefits and refine messaging for efficiency
1 parent ea15dc2 commit 5c8b42d

1 file changed

Lines changed: 5 additions & 9 deletions

File tree

‎README.md‎

Lines changed: 5 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -14,27 +14,23 @@ Context-first AI coding extension for VS Code, Cursor, and others.
1414
❤️ **Free and open-source**—highly capable tooling freely accessible to everyone \
1515
🔒 **Privacy-focused**—runs 100% on your machine, with **zero remote processing** \
1616
🏠️ **Local friendly**—easy on smaller models thanks to **zero tool-calling design** \
17-
🌱 **Efficient**—simple, cacheable prompts, with **zero overhead**
17+
🌱 **Efficient**—largely reused context between tasks enables prompt caching
1818

1919
<p>
2020
<img src="https://github.com/robertpiosik/CodeWebChat/raw/HEAD/packages/shared/src/media/screenshot.png" alt="Screenshot" />
2121
</p>
2222

2323
## Introduction
2424

25-
**LLMs are brilliant, yet mindless pattern matchers.** Because they lack properties similar to mental models, they can't _imagine_ how the correct output should look like _before_ their context is filled with relevant files.
25+
**LLMs are brilliant, yet mindless pattern matchers.** Because they lack properties similar to mental models, they can't _imagine_ how the correct output should look like before their context is filled with relevant files.
2626

27-
To mitigate this limitation, so-called _agentic_ tools ask you to prepare extensive markdown-based plans and feature descriptions, so the system based on a keyword matching (via _grep_ tool calling) can do a better job at automated context compilation. It's a time consuming, token-hungry hit-or-miss.
27+
To mitigate this limitation, other, agentic tools ask you to prepare extensive markdown-based plans and feature descriptions, so the system based on a keyword matching (via _grep_ tool calling) can do a better job at automated context compilation. It's a time consuming, token-hungry hit-or-miss.
2828

29-
**Code Web Chat proposes a mindful 🧠, aligned with true model's capabilities approach**—make the model guide itself with hand-picked files and straightforward instructions, not vague markdowns. **With it comes a big win**—as you work on a task your context becomes largely fixed (recently updated files and your changing instructions are always at the very end), therefore **cache is hit most of the time, making AI coding too cheap to meter!**
30-
31-
**This, combined with the latest generation of models** like Gemini 3.0 Flash and Pro which excel at long context understanding and diff correctness, **marks the new generation of AI coding tools**.
32-
33-
**10X** the speed, **0.1X** the cost, and accuracy that can't get any better.
29+
**Code Web Chat proposes a mindful 🧠, aligned with model's true capabilities approach**—guide the model with context and task-focused instructions. **With it comes a big win**—because all recently updated files and your ever-changing instructions are always at the end of the prompt, **prompt cache is hit for most of its length, making AI coding too cheap to meter!**
3430

3531
### **Context-first approach**
3632

37-
Meet a simple **non-agentic workflow**—select files, enter instructions, and send message with your favorite chatbot or a model provider of choice.
33+
Meet a simple **_non-agentic_ workflow**—select files, enter instructions, and send message with your favorite chatbot or a model provider of choice.
3834

3935
> [!TIP]
4036
> LLMs are pattern matchers—they love examples! Guide the model with context for production-grade code.

0 commit comments

Comments
 (0)