If I work with newer APIs and patterns than the training cutoff date and equip Claude locally with self-invoking skills and knowledge bases, when I do /ultraplan will it use stale knowledge to make the plan?@bcherny@trq212#claudecode
That last 20% is where the real engineering lives — understanding your schema's quirks, your business logic, how your team phrases questions. It's a data engineering problem, not just an AI problem. That's the difference between a toy and a tool people trust.
Type a question, get a chart — looks great on stage. But against a real enterprise database with hundreds of tables, complex relationships, and messy naming conventions? Getting from 80% accuracy to near-100% is where most off-the-shelf approaches fall short.
@tikiverse @bcherny@trq212@AnthropicAI Honestly though, a proper profiles or directory-level config feature would
solve this cleanly. Worth filing at https://t.co/h9Ec9AOvqq if nobody has yet.
Did anyone else notice Claude Code v2.1.81 removed "clear context and accept" from Plan Mode? That was THE recommended workflow. @bcherny@trq212@AnthropicAI what's
the new guidance? #ClaudeCode
@tikiverse @bcherny@trq212@AnthropicAI You can also put shared skills at user scope (~/.claude/) so they're available
everywhere, then just invoke them selectively. And CLAUDE.md can import
shared instruction files across projects.
@tikiverse @bcherny@trq212@AnthropicAI Best workaround: symlinks. Create ~/claude-profiles/coding/ (or /writing/,
/research/) with shared rules/hooks, then symlink into each project's .claude/
directory. Not perfect but avoids the duplication.
@tikiverse @bcherny@trq212@AnthropicAI I am not an expert but here is what i see: Short answer: no, there's nothing between global and per-project settings
right now. It's a real gap.
@JayD0ubleu@bcherny@trq212@AnthropicAI Add "showClearContextOnPlanAccept": true to your settings.json (~/.claude/settings.json for global, or .claude/settings.json for project-level). That restores the 4th option. From the v2.1.81 changelog.
@bcherny@merlindru@trq212@AnthropicAI After plan mode, if CC uses more than 100K tokens i typically clear context and accept the plan just because to reduce the noise ratio. So that is not needed?
@bcherny This is amazing! Thank you! When you run the listerner server and then connect via iOS app, it is always the same session. Am i correct that you cannot end a session from the iOS app and start a new one?
@bcherny@bcherny this is awesome! Would you please explain the cost. Let say i am with $200/m max plan. And for a session i hit 300K tokens. What is the cost for that extra 100K tokes for context please?
When accuracy matters — compliance, financials, specs — "mostly right" isn't good enough. The organizations getting results build purpose-fit solutions with domain knowledge baked in. And enterprise AI APIs today are SOC 2 compliant and don't train on your data.