Skip to content

Add account-state-dependent containers (<user_preferences>, <project_instructions>) and complete Section 20 enumeration#104

Open
SsuHan-Liu wants to merge 2 commits into
asgeirtj:mainfrom
SsuHan-Liu:main
Open

Add account-state-dependent containers (<user_preferences>, <project_instructions>) and complete Section 20 enumeration#104
SsuHan-Liu wants to merge 2 commits into
asgeirtj:mainfrom
SsuHan-Liu:main

Conversation

@SsuHan-Liu
Copy link
Copy Markdown

This PR updates Anthropic/claude-cowork.md with three additions verified via direct introspection on Opus 4.7 Cowork desktop, 2026-04-26.

1. Two account-state-dependent containers (new Sections 21 and 22)

<user_preferences> (snake_case) and <project_instructions> are present in Cowork's system prompt body when the user has Profile preferences or Project Instructions configured. Both are absent when those settings are empty — which likely explains why these containers were not captured in the existing extraction (or in EliFuzz/awesome-system-prompts/leaks/anthropic/2026-01-16_prompt_cowork.md, where grep for user_preferences / userPreferences / project_instructions returns 0 hits).

2. Section 20 enumeration completed

The existing enumeration listed claudeMd + currentDate + Runtime Skills List as the contents of the runtime <system-reminder> block. The block also contains MEMORY.md content and # userEmail. Both are now listed. The closing relevance disclaimer's coverage scope is also clarified (it wraps the entire block, not just any single component).

3. Methodological note

I appended the new sections at the end of the file (Sections 21–22) rather than inserting between current Sections 10 and 11 (where they belong positionally) to keep the diff minimal. Happy to renumber and reposition if preferred — just let me know.

Empirical follow-up

A behavioural A/B study comparing chat (with <preferences_info> meta-rule) and Cowork (without) on Opus 4.7 confirms the structural difference produces visible behaviour change in clean environment, including thinking-trace evidence of Claude applying the meta-rule's case logic at runtime. Full protocol, raw transcripts, and grep evidence: https://github.com/SsuHan-Liu/claude-instruction-containers

Side note (for your awareness, not part of this PR's scope):

The TOC entry 20.1 Runtime Skills List (line 116 of current main) points to a #201-runtime-skills-list anchor that has no matching sub-header in the body. Likely a stale entry. I left it alone since fixing it is outside this PR's scope, but flagging in case you'd like to address separately.

…instructions>) and complete Section 20 enumeration

- New Sections 21–22: <user_preferences> and <project_instructions> containers (verified via direct introspection on Opus 4.7 Cowork desktop, 2026-04-26)
- Section 20: enumeration completed with MEMORY.md and userEmail; closing relevance disclaimer scope clarified

Empirical follow-up: https://github.com/SsuHan-Liu/claude-instruction-containers
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant