Feedback chat justinsforge agent model
name: Justin never runs scripts himself, agent runs them via chat.justinsforge description: Justin corrected the mental model: he doesn't run scripts or CLI tools directly, he uses chat.justinsforge (Session Launcher) to pick project/scenario/account/model and spawns an interactive Claude Code session that runs everything type: feedback
When designing any new tool, skill, or automation, do not spec it as 'Justin runs script X' or 'Justin executes command Y'. Justin's actual workflow is: open chat.justinsforge.com (the Session Launcher, backend forge_session_launcher_api.py:7365), pick project/scenario/modifiers/account/model, which spawns a persistent interactive Claude Code session, then he chats with that agent in natural language and the agent runs scripts/tools on his behalf. Picking a project in the launcher doesn't change cwd, it injects a <session_setup> prompt header naming the repo path (map lives in PROMPT_BLOCKS["targets"]), and the agent cd's/worktrees there itself.
Why: During the CreatorTrack dev-queue design session (2026-07-11), Justin corrected an initial spec that had him running a script directly: 'I don't run scripts I use chat.justin to write prompts and click options and then I chat with claude and it runs scripts.'
How to apply: Any new feature/skill spec should be framed as 'Justin says X in chat, the already-spawned agent invokes tool/skill Y' rather than 'Justin runs command Z'. Reuse PROMPT_BLOCKS["targets"] as the single source of truth for project-to-repo-path mapping instead of building a new config file.
[auto-memory session 5ae6d8f3-bf72-464a-a2e8-61cf3b7368cf, confidence 0.85, mode direct]