How we execute work on this project. These principles guide everything we do here. Captured 2026-10-02 from the owner's directive.
- Always check our assumptions and start with the why.
- Clarify every single assumption we are making — and "clarify" means use research to clear it, not guess.
- We update every assumption at planning and design time, NOT at implementation time. Implementation executes a settled plan; it does not re-open assumptions. Front-load the research (e.g. catching the ElevenLabs streaming limit before building on it) so implementation can run heads-down.
- Only once the assumptions are cleared do we move on to building.
- Once assumptions are clear, do the work using superpowers — always.
- Whenever possible, use a brand-new worktree if we are disrupting a bunch of working code.
- If the size of the work is sizable, parallelize it with a team of sub-agents.
- Run it in goal mode: select the appropriate goal for the work task, then fire it off automatically.
- Whenever a goal is set up, aim for the maximum level of attainment — 10/10 (and if that is too hard, 9/10).
- When an experience touches front-end components (or the front-end is needed to complete the experience), make sure the front-end picks up the change in state and manifests a different view.
- This can be done as a separate pass or at the end of a section — but it is never skipped.
- Keep momentum and keep getting things done.
- Checkpoint the code (commit) whenever a piece builds locally and passes verification.
- Only interrupt the owner for things like API keys (credentials / external enablement). Otherwise, use judgment and proceed.
- Before launching a new section of the codebase or a new task, check the context size first.
- If the context window is above 50%, compact first before starting the new work. (Externalize state to the SDD ledger / git / docs so the resume after compaction is clean.)
Source: owner directive, 2026-10-02.