# Agents: rubrics, skills, references

Three layers, each answering one kind of question. One fact lives in one place; the others cite it by version.

| Layer | Files | Who edits | Loaded by an agent |
|:--|:--|:--|:--|
| Rubrics: fixed, versioned, human-edited | `unit.qmd` (the graph), `modes/*.qmd` (evidence each claim owes), `templates/claim-record.json` | the group, one meeting at a time | on demand, at the node that needs it; `mode-instructions.json` is the machine export of the mode pages, regenerated by `tools/export_modes.py` |
| Skills: lean controllers | `skills/<name>/SKILL.md` | derived from the rubrics after each meeting | in full, when the skill fires |
| References: rationale, sources, examples, evaluation items | `skills/<name>/references/*.md`, `skills/<name>/CHANGELOG.md` | with the rubrics | not by the acting agent; the evaluator loads `evaluation.md`, people load the rest |

Rules for keeping it lean:

- A skill states what to do, what to produce, when to move on, and what not to do. Sources, reasons, and examples go to `references/`.
- A rubric is cited by version, never restated inside a skill or another rubric.
- The evaluator's items are never in the acting agent's context.
- When a meeting changes a rule, edit the rubric first, then the reference, then shorten into the skill, and log the version in the skill's `CHANGELOG.md`.
- Word counts are reported in the changelog; a skill that grows past about a thousand words is split.