- TLH spends tokens to save human attention: isolated developers, reviewers, research, and validation are a feature, not waste.
- The core loop is non-negotiable: think through the work with the architect, approve it, then switch to something else while TLH executes.
- Persistent memory is fundamental: decisions, constraints, and lessons should compound across sessions so settled choices are not ignored.
- Context should be legible to users but managed by TLH, not manually babysat through pruning, compaction, or prompt hygiene.
- TLH stays curated but extensible: every bundled extension, prompt, setting, and workflow must justify its inclusion, while users remain free to add their own.
- TLH is and should be optimised for warm launch time, not cold launch, nor install time.