Simon Willison’s Weblog

Leaked system prompts from Vercel v0. v0 is Vercel's entry in the increasingly crowded LLM-assisted development market - chat with a bot and have that bot build a full application for you.

They've been iterating on it since launching in October last year, making it one of the most mature products in this space.

Somebody leaked the system prompts recently. Vercel CTO Malte Ubl said this:

When @v0 first came out we were paranoid about protecting the prompt with all kinds of pre and post processing complexity.

We completely pivoted to let it rip. A prompt without the evals, models, and especially UX is like getting a broken ASML machine without a manual

# 25th November 2024, 9:17 pm / evals, vercel, ai, llms, prompt-engineering, prompt-injection, ai-assisted-programming, generative-ai

OK, I can partly explain the LLM chess weirdness now (via) Last week Dynomight published Something weird is happening with LLMs and chess pointing out that most LLMs are terrible chess players with the exception of gpt-3.5-turbo-instruct (OpenAI's last remaining completion as opposed to chat model, which they describe as "Similar capabilities as GPT-3 era models").

After diving deep into this, Dynomight now has a theory. It's mainly about completion models v.s. chat models - a completion model like gpt-3.5-turbo-instruct naturally outputs good next-turn suggestions, but something about reformatting that challenge as a chat conversation dramatically reduces the quality of the results.

Through extensive prompt engineering Dynomight got results out of GPT-4o that were almost as good as the 3.5 instruct model. The two tricks that had the biggest impact:

1. Examples. Including just three examples of inputs (with valid chess moves) and expected outputs gave a huge boost in performance.
2. "Regurgitation" - encouraging the model to repeat the entire sequence of previous moves before outputting the next move, as a way to help it reconstruct its context regarding the state of the board.

They experimented a bit with fine-tuning too, but I found their results from prompt engineering more convincing.

No non-OpenAI models have exhibited any talents for chess at all yet. I think that's explained by the A.2 Chess Puzzles section of OpenAI's December 2023 paper Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision:

The GPT-4 pretraining dataset included chess games in the format of move sequence known as Portable Game Notation (PGN). We note that only games with players of Elo 1800 or higher were included in pretraining.

# 21st November 2024, 7:51 pm / prompt-engineering, generative-ai, openai, gpt-4, ai, llms, training-data

ChainForge. I'm still on the hunt for good options for running evaluations against prompts. ChainForge offers an interesting approach, calling itself "an open-source visual programming environment for prompt engineering".

The interface is one of those boxes-and-lines visual programming tools, which reminds me of Yahoo Pipes.

It's open source (from a team at Harvard) and written in Python, which means you can run a local copy instantly via uvx like this:

uvx chainforge serve

You can then configure it with API keys to various providers (OpenAI worked for me, Anthropic models returned JSON parsing errors due to a 500 page from the ChainForge proxy) and start trying it out.

The "Add Node" menu shows the full list of capabilities.

The JavaScript and Python evaluation blocks are particularly interesting: the JavaScript one runs outside of a sandbox using plain eval(), while the Python one still runs in your browser but uses Pyodide in a Web Worker.

# 8th November 2024, 8:52 pm / pyodide, evals, uv, ai, llms, prompt-engineering, prompt-injection, python, javascript, generative-ai

From Naptime to Big Sleep: Using Large Language Models To Catch Vulnerabilities In Real-World Code (via) Google's Project Zero security team used a system based around Gemini 1.5 Pro to find a previously unreported security vulnerability in SQLite (a stack buffer underflow), in time for it to be fixed prior to making it into a release.

A key insight here is that LLMs are well suited for checking for new variants of previously reported vulnerabilities:

A key motivating factor for Naptime and now for Big Sleep has been the continued in-the-wild discovery of exploits for variants of previously found and patched vulnerabilities. As this trend continues, it's clear that fuzzing is not succeeding at catching such variants, and that for attackers, manual variant analysis is a cost-effective approach.

We also feel that this variant-analysis task is a better fit for current LLMs than the more general open-ended vulnerability research problem. By providing a starting point – such as the details of a previously fixed vulnerability – we remove a lot of ambiguity from vulnerability research, and start from a concrete, well-founded theory: "This was a previous bug; there is probably another similar one somewhere".

LLMs are great at pattern matching. It turns out feeding in a pattern describing a prior vulnerability is a great way to identify potential new ones.

# 1st November 2024, 8:15 pm / gemini, security, sqlite, google, generative-ai, ai, llms, prompt-engineering

Generating Descriptive Weather Reports with LLMs. Drew Breunig produces the first example I've seen in the wild of the new LLM attachments Python API. Drew's Downtown San Francisco Weather Vibes project combines output from a JSON weather API with the latest image from a webcam pointed at downtown San Francisco to produce a weather report "with a style somewhere between Jack Kerouac and J. Peterman".

Here's the Python code that constructs and executes the prompt. The code runs in GitHub Actions.

# 29th October 2024, 11:12 pm / vision-llms, drew-breunig, llm, generative-ai, ai, llms, github-actions, prompt-engineering

Introducing Contextual Retrieval (via) Here's an interesting new embedding/RAG technique, described by Anthropic but it should work for any embedding model against any other LLM.

One of the big challenges in implementing semantic search against vector embeddings - often used as part of a RAG system - is creating "chunks" of documents that are most likely to semantically match queries from users.

Anthropic provide this solid example where semantic chunks might let you down:

Imagine you had a collection of financial information (say, U.S. SEC filings) embedded in your knowledge base, and you received the following question: "What was the revenue growth for ACME Corp in Q2 2023?"

A relevant chunk might contain the text: "The company's revenue grew by 3% over the previous quarter." However, this chunk on its own doesn't specify which company it's referring to or the relevant time period, making it difficult to retrieve the right information or use the information effectively.

Their proposed solution is to take each chunk at indexing time and expand it using an LLM - so the above sentence would become this instead:

This chunk is from an SEC filing on ACME corp's performance in Q2 2023; the previous quarter's revenue was $314 million. The company's revenue grew by 3% over the previous quarter."

This chunk was created by Claude 3 Haiku (their least expensive model) using the following prompt template:

<document>
{{WHOLE_DOCUMENT}}
</document>
Here is the chunk we want to situate within the whole document
<chunk>
{{CHUNK_CONTENT}}
</chunk>
Please give a short succinct context to situate this chunk within the overall document for the purposes of improving search retrieval of the chunk. Answer only with the succinct context and nothing else.

Here's the really clever bit: running the above prompt for every chunk in a document could get really expensive thanks to the inclusion of the entire document in each prompt. Claude added context caching last month, which allows you to pay around 1/10th of the cost for tokens cached up to your specified beakpoint.

By Anthropic's calculations:

Assuming 800 token chunks, 8k token documents, 50 token context instructions, and 100 tokens of context per chunk, the one-time cost to generate contextualized chunks is $1.02 per million document tokens.

Anthropic provide a detailed notebook demonstrating an implementation of this pattern. Their eventual solution combines cosine similarity and BM25 indexing, uses embeddings from Voyage AI and adds a reranking step powered by Cohere.

The notebook also includes an evaluation set using JSONL - here's that evaluation data in Datasette Lite.

# 20th September 2024, 1:34 am / anthropic, claude, generative-ai, ai, embeddings, llms, search, rag, prompt-engineering

files-to-prompt 0.3. New version of my files-to-prompt CLI tool for turning a bunch of files into a prompt suitable for piping to an LLM, described here previously.

It now has a -c/--cxml flag for outputting the files in Claude XML-ish notation (XML-ish because it's not actually valid XML) using the format Anthropic describe as recommended for long context:

files-to-prompt llm-*/README.md --cxml | llm -m claude-3.5-sonnet \
  --system 'return an HTML page about these plugins with usage examples' \
  > /tmp/fancy.html

Here's what that gave me.

The format itself looks something like this:

<documents>
<document index="1">
<source>llm-anyscale-endpoints/README.md</source>
<document_content>
# llm-anyscale-endpoints
...
</document_content>
</document>
</documents>

# 9th September 2024, 5:57 am / anthropic, claude, tools, projects, generative-ai, ai, llms, prompt-engineering

llm-claude-3 0.4.1. New minor release of my LLM plugin that provides access to the Claude 3 family of models. Claude 3.5 Sonnet recently upgraded to a 8,192 output limit recently (up from 4,096 for the Claude 3 family of models). LLM can now respect that.

The hardest part of building this was convincing Claude to return a long enough response to prove that it worked. At one point I got into an argument with it, which resulted in this fascinating hallucination:

I eventually got a 6,162 token output using:

cat long.txt | llm -m claude-3.5-sonnet-long --system 'translate this document into french, then translate the french version into spanish, then translate the spanish version back to english. actually output the translations one by one, and be sure to do the FULL document, every paragraph should be translated correctly. Seriously, do the full translations - absolutely no summaries!'

# 30th August 2024, 11:28 pm / llm, anthropic, claude, generative-ai, projects, ai, llms, prompt-engineering, claude-3-5-sonnet

Anthropic’s Prompt Engineering Interactive Tutorial (via) Anthropic continue their trend of offering the best documentation of any of the leading LLM vendors. This tutorial is delivered as a set of Jupyter notebooks - I used it as an excuse to try uvx like this:

git clone https://github.com/anthropics/courses
uvx --from jupyter-core jupyter notebook courses

This installed a working Jupyter system, started the server and launched my browser within a few seconds.

The first few chapters are pretty basic, demonstrating simple prompts run through the Anthropic API. I used %pip install anthropic instead of !pip install anthropic to make sure the package was installed in the correct virtual environment, then filed an issue and a PR.

One new-to-me trick: in the first chapter the tutorial suggests running this:

API_KEY = "your_api_key_here"
%store API_KEY

This stashes your Anthropic API key in the IPython store. In subsequent notebooks you can restore the API_KEY variable like this:

%store -r API_KEY

I poked around and on macOS those variables are stored in files of the same name in ~/.ipython/profile_default/db/autorestore.

Chapter 4: Separating Data and Instructions included some interesting notes on Claude's support for content wrapped in XML-tag-style delimiters:

Note: While Claude can recognize and work with a wide range of separators and delimeters, we recommend that you use specifically XML tags as separators for Claude, as Claude was trained specifically to recognize XML tags as a prompt organizing mechanism. Outside of function calling, there are no special sauce XML tags that Claude has been trained on that you should use to maximally boost your performance. We have purposefully made Claude very malleable and customizable this way.

Plus this note on the importance of avoiding typos, with a nod back to the problem of sandbagging where models match their intelligence and tone to that of their prompts:

This is an important lesson about prompting: small details matter! It's always worth it to scrub your prompts for typos and grammatical errors. Claude is sensitive to patterns (in its early years, before finetuning, it was a raw text-prediction tool), and it's more likely to make mistakes when you make mistakes, smarter when you sound smart, sillier when you sound silly, and so on.

Chapter 5: Formatting Output and Speaking for Claude includes notes on one of Claude's most interesting features: prefill, where you can tell it how to start its response:

client.messages.create(
    model="claude-3-haiku-20240307",
    max_tokens=100,
    messages=[
        {"role": "user", "content": "JSON facts about cats"},
        {"role": "assistant", "content": "{"}
    ]
)

Things start to get really interesting in Chapter 6: Precognition (Thinking Step by Step), which suggests using XML tags to help the model consider different arguments prior to generating a final answer:

Is this review sentiment positive or negative? First, write the best arguments for each side in <positive-argument> and <negative-argument> XML tags, then answer.

The tags make it easy to strip out the "thinking out loud" portions of the response.

It also warns about Claude's sensitivity to ordering. If you give Claude two options (e.g. for sentiment analysis):

In most situations (but not all, confusingly enough), Claude is more likely to choose the second of two options, possibly because in its training data from the web, second options were more likely to be correct.

This effect can be reduced using the thinking out loud / brainstorming prompting techniques.

A related tip is proposed in Chapter 8: Avoiding Hallucinations:

How do we fix this? Well, a great way to reduce hallucinations on long documents is to make Claude gather evidence first.

In this case, we tell Claude to first extract relevant quotes, then base its answer on those quotes. Telling Claude to do so here makes it correctly notice that the quote does not answer the question.

I really like the example prompt they provide here, for answering complex questions against a long document:

<question>What was Matterport's subscriber base on the precise date of May 31, 2020?</question>

Please read the below document. Then, in <scratchpad> tags, pull the most relevant quote from the document and consider whether it answers the user's question or whether it lacks sufficient detail. Then write a brief numerical answer in <answer> tags.

# 30th August 2024, 2:52 am / anthropic, claude, uv, ai, llms, prompt-engineering, python, generative-ai, jupyter

System prompt for val.town/townie (via) Val Town (previously) provides hosting and a web-based coding environment for Vals - snippets of JavaScript/TypeScript that can run server-side as scripts, on a schedule or hosting a web service.

Townie is Val's new AI bot, providing a conversational chat interface for creating fullstack web apps (with blob or SQLite persistence) as Vals.

In the most recent release of Townie Val added the ability to inspect and edit its system prompt!

I've archived a copy in this Gist, as a snapshot of how Townie works today. It's surprisingly short, relying heavily on the model's existing knowledge of Deno and TypeScript.

I enjoyed the use of "tastefully" in this bit:

Tastefully add a view source link back to the user's val if there's a natural spot for it and it fits in the context of what they're building. You can generate the val source url via import.meta.url.replace("esm.town", "val.town").

The prompt includes a few code samples, like this one demonstrating how to use Val's SQLite package:

import { sqlite } from "https://esm.town/v/stevekrouse/sqlite";
let KEY = new URL(import.meta.url).pathname.split("/").at(-1);
(await sqlite.execute(`select * from ${KEY}_users where id = ?`, [1])).rows[0].id

It also reveals the existence of Val's very own delightfully simple image generation endpoint Val, currently powered by Stable Diffusion XL Lightning on fal.ai.

If you want an AI generated image, use https://maxm-imggenurl.web.val.run/the-description-of-your-image to dynamically generate one.

Here's a fun colorful raccoon with a wildly inappropriate hat.

Val are also running their own gpt-4o-mini proxy, free to users of their platform:

import { OpenAI } from "https://esm.town/v/std/openai";
const openai = new OpenAI();
const completion = await openai.chat.completions.create({
  messages: [
    { role: "user", content: "Say hello in a creative way" },
  ],
  model: "gpt-4o-mini",
  max_tokens: 30,
});

Val developer JP Posma wrote a lot more about Townie in How we built Townie – an app that generates fullstack apps, describing their prototyping process and revealing that the current model it's using is Claude 3.5 Sonnet.

Their current system prompt was refined over many different versions - initially they were including 50 example Vals at quite a high token cost, but they were able to reduce that down to the linked system prompt which includes condensed documentation and just one templated example.

# 28th August 2024, 3:33 am / claude-3-5-sonnet, deno, sqlite, anthropic, claude, typescript, ai, llms, prompt-engineering, ai-assisted-programming, javascript, generative-ai, val-town, text-to-image

Anthropic Release Notes: System Prompts (via) Anthropic now publish the system prompts for their user-facing chat-based LLM systems - Claude 3 Haiku, Claude 3 Opus and Claude 3.5 Sonnet - as part of their documentation, with a promise to update this to reflect future changes.

Currently covers just the initial release of the prompts, each of which is dated July 12th 2024.

Anthropic researcher Amanda Askell broke down their system prompt in detail back in March 2024. These new releases are a much appreciated extension of that transparency.

These prompts are always fascinating to read, because they can act a little bit like documentation that the providers never thought to publish elsewhere.

There are lots of interesting details in the Claude 3.5 Sonnet system prompt. Here's how they handle controversial topics:

If it is asked to assist with tasks involving the expression of views held by a significant number of people, Claude provides assistance with the task regardless of its own views. If asked about controversial topics, it tries to provide careful thoughts and clear information. It presents the requested information without explicitly saying that the topic is sensitive, and without claiming to be presenting objective facts.

Here's chain of thought "think step by step" processing baked into the system prompt itself:

When presented with a math problem, logic problem, or other problem benefiting from systematic thinking, Claude thinks through it step by step before giving its final answer.

Claude's face blindness is also part of the prompt, which makes me wonder if the API-accessed models might more capable of working with faces than I had previously thought:

Claude always responds as if it is completely face blind. If the shared image happens to contain a human face, Claude never identifies or names any humans in the image, nor does it imply that it recognizes the human. [...] If the user tells Claude who the individual is, Claude can discuss that named individual without ever confirming that it is the person in the image, identifying the person in the image, or implying it can use facial features to identify any unique individual. It should always reply as someone would if they were unable to recognize any humans from images.

It's always fun to see parts of these prompts that clearly hint at annoying behavior in the base model that they've tried to correct!

Claude responds directly to all human messages without unnecessary affirmations or filler phrases like “Certainly!”, “Of course!”, “Absolutely!”, “Great!”, “Sure!”, etc. Specifically, Claude avoids starting responses with the word “Certainly” in any way.

Anthropic note that these prompts are for their user-facing products only - they aren't used by the Claude models when accessed via their API.

# 26th August 2024, 8:05 pm / prompt-engineering, anthropic, claude, generative-ai, ai, llms

Long context prompting tips (via) Interesting tips here from Anthropic's documentation about how to best prompt Claude to work with longer documents.

Put longform data at the top: Place your long documents and inputs (~20K+ tokens) near the top of your prompt, above your query, instructions, and examples. This can significantly improve Claude’s performance across all models. Queries at the end can improve response quality by up to 30% in tests, especially with complex, multi-document inputs.

It recommends using not-quite-valid-XML to add those documents to those prompts, and using a prompt that asks Claude to extract direct quotes before replying to help it focus its attention on the most relevant information:

Find quotes from the patient records and appointment history that are relevant to diagnosing the patient's reported symptoms. Place these in <quotes> tags. Then, based on these quotes, list all information that would help the doctor diagnose the patient's symptoms. Place your diagnostic information in <info> tags.

# 26th August 2024, 6:39 pm / prompt-engineering, anthropic, claude, generative-ai, ai, llms

AI-powered Git Commit Function (via) Andrej Karpathy built a shell alias, gcm, which passes your staged Git changes to an LLM via my LLM tool, generates a short commit message and then asks you if you want to "(a)ccept, (e)dit, (r)egenerate, or (c)ancel?".

Here's the incantation he's using to generate that commit message:

git diff --cached | llm "
Below is a diff of all staged changes, coming from the command:
\`\`\`
git diff --cached
\`\`\`
Please generate a concise, one-line commit message for these changes."

This pipes the data into LLM (using the default model, currently gpt-4o-mini unless you set it to something else) and then appends the prompt telling it what to do with that input.

# 26th August 2024, 1:06 am / llm, ai, llms, andrej-karpathy, prompt-engineering, ai-assisted-programming, generative-ai, git

LLMs are bad at returning code in JSON (via) Paul Gauthier's Aider is a terminal-based coding assistant which works against multiple different models. As part of developing the project Paul runs extensive benchmarks, and his latest shows an interesting result: LLMs are slightly less reliable at producing working code if you request that code be returned as part of a JSON response.

The May release of GPT-4o is the closest to a perfect score - the August appears to have regressed slightly, and the new structured output mode doesn't help and could even make things worse (though that difference may not be statistically significant).

Paul recommends using Markdown delimiters here instead, which are less likely to introduce confusing nested quoting issues.

# 16th August 2024, 5:04 pm / prompt-engineering, generative-ai, json, ai, llms, aider, evals

Prompt caching with Claude (via) The Claude API now supports prompt caching, allowing you to mark reused portions of long prompts (like a large document provided as context). Claude will cache these for up to five minutes, and any prompts within that five minutes that reuse the context will be both significantly faster and will be charged at a significant discount: ~10% of the cost of sending those uncached tokens.

Writing to the cache costs money. The cache TTL is reset every time it gets a cache hit, so any application running more than one prompt every five minutes should see significant price decreases from this. If you app prompts less than once every five minutes you'll be losing money.

This is similar to Google Gemini's context caching feature, but the pricing model works differently. Gemini charge $4.50/million tokens/hour for their caching (that's for Gemini 1.5 Pro - Gemini 1.5 Flash is $1/million/hour), for a quarter price discount on input tokens (see their pricing).

Claude’s implementation also appears designed to help with ongoing conversations. Using caching during an individual user’s multi-turn conversation - where a full copy of the entire transcript is sent with each new prompt - could help even for very low traffic (or even single user) applications.

Here's the full documentation for the new Claude caching feature, currently only enabled if you pass "anthropic-beta: prompt-caching-2024-07-31" as an HTTP header.

Interesting to note that this caching implementation doesn't save on HTTP overhead: if you have 1MB of context you still need to send a 1MB HTTP request for every call. I guess the overhead of that HTTP traffic is negligible compared to the overhead of processing those tokens once they arrive.

One minor annoyance in the announcement for this feature:

Detailed instruction sets: Share extensive lists of instructions, procedures, and examples to fine-tune Claude's responses. [...]

I wish Anthropic wouldn't use the term "fine-tune" in this context (they do the same thing in their tweet). This feature is unrelated to model fine-tuning (a feature Claude provides via AWS Bedrock). People find this terminology confusing already, frequently misinterpreting "fine-tuning" as being the same thing as "tweaking your prompt until it works better", and Anthropic's language here doesn't help.

# 14th August 2024, 5:07 pm / gemini, prompt-engineering, anthropic, claude, generative-ai, ai, llms, llm-pricing

Using gpt-4o-mini as a reranker. Tip from David Zhang: "using gpt-4-mini as a reranker gives you better results, and now with strict mode it's just as reliable as any other reranker model".

David's code here demonstrates the Vercel AI SDK for TypeScript, and its support for structured data using Zod schemas.

const res = await generateObject({
  model: gpt4MiniModel,
  prompt: `Given the list of search results, produce an array of scores measuring the liklihood of the search result containing information that would be useful for a report on the following objective: ${objective}\n\nHere are the search results:\n<results>\n${resultsString}\n</results>`,
  system: systemMessage(),
  schema: z.object({
    scores: z
      .object({
        reason: z
          .string()
          .describe(
            'Think step by step, describe your reasoning for choosing this score.',
          ),
        id: z.string().describe('The id of the search result.'),
        score: z
          .enum(['low', 'medium', 'high'])
          .describe(
            'Score of relevancy of the result, should be low, medium, or high.',
          ),
      })
      .array()
      .describe(
        'An array of scores. Make sure to give a score to all ${results.length} results.',
      ),
  }),
});

It's using the trick where you request a reason key prior to the score, in order to implement chain-of-thought - see also Matt Webb's Braggoscope Prompts.

# 11th August 2024, 6:06 pm / gpt-4, llms, ai, generative-ai, prompt-engineering

Braggoscope Prompts. Matt Webb's Braggoscope (previously) is an alternative way to browse the archive's of the BBC's long-running radio series In Our Time, including the ability to browse by Dewey Decimal library classification, view related episodes and more.

Matt used an LLM to generate the structured data for the site, based on the episode synopsis on the BBC's episode pages like this one.

The prompts he used for this are now described on this new page on the site.

Of particular interest is the way the Dewey Decimal classifications are derived. Quoting an extract from the prompt:

- Provide a Dewey Decimal Classification code, label, and reason for the classification.

- Reason: summarise your deduction process for the Dewey code, for example considering the topic and era of history by referencing lines in the episode description. Bias towards the main topic of the episode which is at the beginning of the description.

- Code: be as specific as possible with the code, aiming to give a second level code (e.g. "510") or even lower level (e.g. "510.1"). If you cannot be more specific than the first level (e.g. "500"), then use that.

Return valid JSON conforming to the following Typescript type definition:

{
    "dewey_decimal": {"reason": string, "code": string, "label": string}
}

That "reason" key is essential, even though it's not actually used in the resulting project. Matt explains why:

It gives the AI a chance to generate tokens to narrow down the possibility space of the code and label that follow (the reasoning has to appear before the Dewey code itself is generated).

Here's a relevant note from OpenAI's new structured outputs documentation:

When using Structured Outputs, outputs will be produced in the same order as the ordering of keys in the schema.

That's despite JSON usually treating key order as undefined. I think OpenAI designed the feature to work this way precisely to support the kind of trick Matt is using for his Dewey Decimal extraction process.

# 7th August 2024, 11:23 pm / matt-webb, prompt-engineering, generative-ai, ai, llms, structured-extraction