persistlisted

## Core idea - Context window = RAM (volatile, limited) - Anything important gets written to disk. ## Workflow ### Step 1: Determine task type Before doing any work, determine whether the user's request is primarily about **understanding or documenting existing code** in the current repository. - **If yes** → invoke the `research-codebase` skill instead. Do not continue with this skill. - Examples: "how does the auth flow work?", "document the payment service", "explain the middleware chain", "what classes handle X?" - **If no** → continue with Step 2. - Examples: no argument or "analyse this RFC", "summarise the compliance requirements", "investigate options for a new library", "compare approaches for X" ### Step 2: Resolve the argument **If an argument was provided:** - Treat it as the topic/subject to persist — use it to derive `{topic}` and proceed to Step 3. **If no argument was provided:** - Summarise the conversation since the last persistence event (i.e. since the last invocation of `persist`, `planning`, `planning-mode`, `research-mode`, or any other skill that writes to `.ai/`). If no such event exists, summarise the entire conversation. - The summary must be structured around **learnings, insights, and decisions** — not a chronological account of what was said. Extract the substance: what was established, what was ruled out, what trade-offs were identified, what conclusions were reached. - Derive `{topic}` from the subject matter of that conversation