Daily digest
AI digest: 2026-11-16
Highlights
- Claude Code Code Mode can now be tried: the docs page is live. Start Claude Code
with
CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1(interactive terminal only, 5.5 models or Fable 5.2) and Claude writes JavaScript that makes several tool calls at once. - Claude Code 2.1.325 changes permission defaults: sandboxed Bash prompts again
unless
sandbox.autoAllowBashIfSandboxed: true, and scheduled tasks or/loopneed aSkill(name)allow rule to run skills that declare tools or hooks. - craft: a new design-process skill pack (brief, every screen state, critique):
/plugin marketplace add diego3g/craft, then/plugin install craft@craft.
New sources
Agents and tooling
- Run tool calls from code with Code Mode: how to turn on Claude Code's experimental Code Mode, its requirements, limits, and permission checks.
- Claude Code release 2.1.325: new permission defaults for sandboxed Bash, skills in scheduled tasks, and git outside the project; many security fixes.
- craft: design skills for AI agents: skills for design briefs, screen states, directions, and critique; Claude Code, Codex, Cursor.
- Update docx, pptx, xlsx, and pdf skills: stricter validation, new pptx helpers, and an xlsx recalculation fix in Anthropic's document skills.
- XERJ: local Rust search engine over code, docs, and logs with an MCP server.
Research
- Building strong safety evaluations: Anthropic's checklist for eval design, grader validation, sample sizes, and reporting.
Topic pages updated
- tooling/coding-agents: Code Mode setup and limits; 2.1.325 permission defaults; Code Mode open question answered.
- practices/security: 2.1.325 security fixes and tighter defaults.
- tooling/skills: craft, document skills update, skills in scheduled tasks.
- tooling/mcp: XERJ local search server.
- agents/tool-use: new section on tools called from code.
- practices/evals: validity, grader validation, sample sizes, and configuration logging.