⚙️System Prompt/dynamic
Session-Specific Guidance (documentation summary)
src/constants/prompts.ts
Prompt Engineering Insight
Educational digest of the dynamic session-guidance registry: it gates behavior on which tools exist so the model does not hallucinate capabilities. Compare system-discover-skills for the discover-tool slice.
Techniques Used
conditional-logictool-use-guidancetaxonomyguardrails
prompt
Session-specific guidance
The live system prompt builds this section from many conditional fragments (tool availability, fork mode, skills, verification, etc.). This entry summarizes the themes; exact bullets and tool names are interpolated at runtime.
- If AskUserQuestion is available: use it when you do not understand why the user denied a tool call.
- Interactive sessions: users may run shell commands in-session with
! <command>so output lands in the conversation (e.g. gcloud auth login). - Agent tool: fork mode explains background forks; otherwise use specialized subagents for parallel work without duplicating delegated research.
- Explore agent: for simple directed search use Glob/Grep (or bash find/grep when embedded); use Explore subagent only for broad exploration (threshold applies).
- Skills: /skill-name shorthand expands to prompts; invoke only via Skill tool for user-invocable skills listed there — do not guess built-in CLI commands.
- DiscoverSkills (experimental): when skill search is enabled, call the discover tool for pivots not covered by surfaced skills.
- Verification agent (gated): non-trivial implementation requires adversarial verification subagent before reporting completion; verifier-owned PASS/FAIL — spot-check commands from report.
Tags
sessiontoolsskillsverification