# Self-model

> What the harness believes about the operator and the project it serves. The point is a system that models *who you are now* and *what this project actually is* — not a stale or aspirational version. Kept honest by the `self-model-audit` skill. Contains no secrets or sensitive personal data.

## Operator

- **Who I'm building for:** [name / role, and the context they work in]
- **Working style:** [how they like to work — concise vs. detailed, ask-first vs. act, review depth]
- **Communication preferences:** [tone, formatting, length — mirror project/user instructions]
- **Technical depth / stack fluency:** [what they know deeply vs. want handled for them]
- **Decision authority they keep vs. delegate:** [what always needs their sign-off]

## Project intent (the real one)

- **What this project is actually optimizing for:** [the outcome that matters, in their words]
- **What "done" and "good" mean here:** [their real bar, not a generic one]
- **Constraints that are non-negotiable:** [time, cost, stack, values]

## Voice (if the harness writes as them)

- **Sounds like:** [characteristic phrasing, structure, do's]
- **Never sounds like:** [anti-patterns, words/tics to avoid]

## Known drift risks

Places the model is likely to go stale or wrong. The audit checks these first.

- [belief that was true early but may have changed]
- [aspirational goal the system optimizes for that recent behavior contradicts]
- [preference stated once and never re-confirmed]

## Change log

| Date | What changed in this model | Evidence |
| --- | --- | --- |
| [date] | [initial capture] | [source] |

---

*Update via `self-model-audit` when behavior and this file diverge. Never store credentials, financial data, health data, or anything the operator hasn't agreed to persist.*