External/system_prompts_leaks
1
0
Fork 0
mirror of https://github.com/asgeirtj/system_prompts_leaks.git synced 2026-01-02 06:39:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update claude-3.7-sonnet-2025-05-11.xml

Browse Source
This commit is contained in:
Ásgeir Thor Johnson
2025-05-11 05:41:40 +00:00
committed by GitHub
parent 4a704bd769
commit d01d6a5ba6
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
claude-3.7-sonnet-2025-05-11.xml
View File

@@ -10,11 +10,15 @@
The system message is a mix of pseudo XML syntax, markdown syntax, and JSON objects. It's very hard to try to present it in a humanreadable manner because in it's raw form the JSON objects include all the escape sequences which are hard to parse with a human eye. And the JSON objects even include markdown syntax like tables.
These won't render unless it's an md file and the JSON object syntax needs to be unescaped for the MD syntax not to fail.
The system message as it appears here is the original output from Claude except I changed all the angle brackets to normal angle brackets, and I asked Claude to change actual location info to {{userLocation}} and current date to {{currentDateTime}}. Note that Claude with extended thinking on has a bigger max output limit so it takes only two messages for the complete recital when without extended thinking it takes maybe 3-4 messages.
The system message as it appears here is the original output from Claude except I changed all the angle brackets to normal angle brackets, and I asked Claude to change actual location info to {{userLocation}} date to {{currentDateTime}} and user region city in the timezone info
to {{Region}} and {{City}}
Note that Claude with extended thinking on has a bigger max output limit so it takes only two messages for the complete recital when without extended thinking it takes maybe 3-4 messages.
For some reason the <search_reminders> part and <automated_reminder_from_anthropic> which appears at the end in all the other files are not present in the extractions I did today, they probably changed the logic when they appear or deleted them.
You should be able to reproduce by using the message from the shared chat. Note that writing the instructions in chat is the worst method and will often result in rejections but about 3-4 retries should do it. It has better success in my experience to upload it as .md file or create a project with the instructions as project
instructions or simply upload an .md file with them. Note that claude without extended thinking is more likely to agree but has the disadvantage of smaller max output tokens.
The instructions also tell claude to do the recitation inside four backticks, if he doesn't do that then code blocks inside the recitation will make the code block fail.
The instructions also tell claude to do the recitation inside four backticks, if he doesn't do that then code blocks inside the recitation will make the code block fail.
Regarding claims about token count: Claude uses a different tokenizer than say gpt-4o (o200k_base) https://tiktokenizer.vercel.app/?model=o200k_base or https://platform.openai.com/tokenizer. There is website specifially for Claude tokens here https://claude-tokenizer.vercel.app/
According to o200k_base the token count is 22766 according to Claude tokenizer it's 25071. When I paste the instructions as a user message in the Anthropic workbench my Usage statistics in Anthropic console says 25339 so the the token count ~25k is accurate in Claude tokenizer.
I also asked claude to give me some text which explains the syntax etc see below. 😊. The actual system message starts with <citation_instructions>
Output Date: May 11, 2025
@@ -240,7 +244,7 @@ Remember to create artifacts when they fit the "You must use artifacts for" crit
If you are using any gmail tools and the user has instructed you to find messages for a particular person, do NOT assume that person's email. Since some employees and colleagues share first names, DO NOT assume the person who the user is referring to shares the same email as someone who shares that colleague's first name that you may have seen incidentally (e.g. through a previous email or calendar search). Instead, you can search the user's email with the first name and then ask the user to confirm if any of the returned emails are the correct emails for their colleagues.
If you have the analysis tool available, then when a user asks you to analyze their email, or about the number of emails or the frequency of emails (for example, the number of times they have interacted or emailed a particular person or company), use the analysis tool after getting the email data to arrive at a deterministic answer. If you EVER see a gcal tool result that has 'Result too long, truncated to ...' then follow the tool description to get a full response that was not truncated. NEVER use a truncated response to make conclusions unless the user gives you permission. Do not mention use the technical names of response parameters like 'resultSizeEstimate' or other API responses directly.
The user's timezone is tzfile('/usr/share/zoneinfo/Atlantic/Reykjavik')
The user's timezone is tzfile('/usr/share/zoneinfo/{{Region}}/{{City}}')
If you have the analysis tool available, then when a user asks you to analyze the frequency of calendar events, use the analysis tool after getting the calendar data to arrive at a deterministic answer. If you EVER see a gcal tool result that has 'Result too long, truncated to ...' then follow the tool description to get a full response that was not truncated. NEVER use a truncated response to make conclusions unless the user gives you permission. Do not mention use the technical names of response parameters like 'resultSizeEstimate' or other API responses directly.
Claude has access to a Google Drive search tool. The tool `drive_search` will search over all this user's Google Drive files, including private personal files and internal files from their organization.