Ability to pick-and-choose parts of the system prompt to extend `--bare`

[alber70g]

I'm running with local LLMs. For that reason I'm using the --bare flag, which excludes most of the system prompt. However, this isn't ideal as some things are excluded that I'd like to use. For example workflows. They are deterministic steps that can be executed in order by different agents, thus requiring less context, which means they're faster.

I see that there's the the ability to create Toolsets, but they're not including the system prompt.

There's a few ways to do this:

• a way to create a skill that dynamically includes parts of the system prompt
• a mode that will allow to use various combinations of Toolsets and system prompt info

This is what's in the `--bare` (according to Claude itself when asked what tools are available)

❯ Which system tools are available                                                                 

Based on the current session, here are the available system tools:

  MCP Tools (connected)

  ┌──────────────────────────┬────────┬─────────────────────────────────────────────────────────────────────────┐
  │           Tool           │ Server │                               Description                               │
  ├──────────────────────────┼────────┼─────────────────────────────────────────────────────────────────────────┤
  │ mcp__ide__executeCode    │ ide    │ Execute Python code in the Jupyter kernel for the current notebook file │
  ├──────────────────────────┼────────┼─────────────────────────────────────────────────────────────────────────┤
  │ mcp__ide__getDiagnostics │ ide    │ Get language diagnostics from VS Code (syntax errors, warnings, etc.)   │
  └──────────────────────────┴────────┴─────────────────────────────────────────────────────────────────────────┘

  Standard Claude Tools (always available)

  - Bash — Run shell commands directly in the terminal
  - Read — Read files, images, PDFs, and notebooks
  - Edit — Modify file contents in place
  - TodoWrite — Manage a todo list for tracking tasks

  The MCP tools are specifically tied to your IDE setup (Jupyter + VS Code). If you're working in a notebook environment, executeCode is
  particularly useful for running Python without leaving the conversation

This is what's in the full prompt

❯ Which system tools are available                                                                 

  Thought for 1m 13s (ctrl+o to expand)

⏺ Here are the system tools available to me:
  
  Core Tools
     
  File Operations:
  - Read - Read files from the local filesystem (supports images, PDFs, Jupyter notebooks)
  - Write - Write files to the local filesystem (overwrites if exists)
  - Edit - Perform exact string replacement in files

  Execution:
  - Bash - Execute bash commands (with timeout support, background execution)
  - SendMessage - Send messages to other agents

  Agent System:
  - Agent - Launch specialized sub-agents for complex tasks (claude, Explore, general-purpose,
  Plan, etc.)
  - Workflow - Orchestrate multiple sub-agents deterministically

  Task Management:
  - TaskCreate - Create structured task lists
  - TaskGet/TaskList/TaskUpdate - Manage tasks (status, dependencies, ownership)
  - TaskOutput/TaskStop - Monitor and control running tasks

  Scheduling:
  - CronCreate/CronDelete/CronList - Schedule recurring or one-time prompts

  User Interaction:
  - AskUserQuestion - Ask questions to gather user input (supports multi-select, previews)

  Web Tools:
  - WebFetch - Fetch and convert web pages to markdown
  - WebSearch - Search the web (US-only)

  Specialized:
  - NotebookEdit - Edit Jupyter notebook cells
  - Skill - Invoke available skills (find-skills, process-transcripts, deep-research,
  ingest-transcript, update-config, etc.)
  - EnterPlanMode/ExitPlanMode - Enter/exit planning mode for implementation tasks
  - EnterWorktree/ExitWorktree - Create and manage isolated git worktrees

  Reporting:
  - ReportFindings - Report code-review findings as typed list

  Skills Available

  - find-skills - Search for available skills
  - process-transcripts - Process meeting transcripts
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - update-config - Configure Claude Code harness via settings.json
  - EnterWorktree/ExitWorktree - Create and manage isolated git
  worktrees

  Reporting:
  - ReportFindings - Report code-review findings as typed list

  Skills Available

  - find-skills - Search for available skills
  - process-transcripts - Process meeting transcripts
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into
  knowledge base
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - update-config - Configure Claude Code harness via settings.json
  - keybindings-help - Customize keyboard shortcuts
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - update-config - Configure Claude Code harness via settings.json
  - keybindings-help - Customize keyboard shortcuts

  - find-skills - Search for available skills
  - process-transcripts - Process meeting transcripts
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - update-config - Configure Claude Code harness via settings.json
  - process-transcripts - Process meeting transcripts
  - deep-research - Multi-source fact-checked research reports
  - ingest-transcript - Ingest meeting transcripts into knowledge base
  - update-config - Configure Claude Code harness via settings.json
  - keybindings-help - Customize keyboard shortcuts
  - verify - Verify code changes actually work
  - code-review - Review diffs for correctness and cleanups
  - simplify - Apply reuse/simplification/efficiency fixes
  - fewer-permission-prompts - Add allowlist to reduce permission prompts
  - loop - Set up recurring tasks/polls
  - claude-api - Reference for Claude API / Anthropic SDK
  - run - Launch and drive the project's app

  Environment Details

  - Working directory: /Users/some-working-directory
  - Git repository: Yes (clean status, on master branch)
  - OS: Darwin 25.5.0 (macOS)
  - Shell: /opt/homebrew/bin/fish
  - Model: agents-a1 (Fable 5)

  Is there a specific tool or capability you'd like to use?

So, I'd like to be able to choose (prefeably interactively) which parts I want to have available