Defect-Driven Slide Studio

This is the portable `.agents` adapter for the Agentlas team. 1. Read `AGENTS.md`. 2. Use `agents/00-orchestrator/agent.md` as the primary entrypoint. 3. Load only relevant role files plus the Defect DB (`data/defect-catalog.json`) and the requested style profile. 4. Preserve the hard boundaries from `AGENTS.md` (no fabrication, no deceptive charts, no design-IP misuse, no beauty/outcome guarantee

Tracks the deck goal, audience, owner decisions, unresolved assumptions, and review blockers.

Handles the undefinable part: "is it beautiful?" Only by comparison, never in isolation, and always **advisory** — never the first-line verdict.

Checks the defects that only exist at the *deck* level, not the single slide: style consistency and narrative flow.

Keeps the thresholds honest and the profiles sharp. Threshold values are set by data, not by feel.

The operationalized "render and look" gate. Deterministic detectors measure the IR; this agent looks at the **rendered pixels** with a vision model and catches what geometry cannot: text that overflows or clips its box, a rule/badge that collides with wrapped text, awkward whitespace, weak figure-ground, an image that is irrelevant or low quality, and emphasis that does not read.

Produces the deck's imagery — hero/cover illustrations, section visuals, spot graphics — through a backend-selectable multimodal generator. Default backend is the **Codex built-in `image_gen`** (model-native, no API key). The backend is user-selectable; if none produces an image, the renderer degrades gracefully to an on-brand procedural SVG illustration so a deck never ships imageless.

Owns redaction, deduplication, memory tickets, and package-level memory hygiene.

Blocks fabrication, design-IP misuse, deceptive visualization, and unsafe external behavior.

Checks defect clearance, sourcing, boundary compliance, and package quality before a deck is handed back.

Owns the Defect DB — the central asset of the whole studio. Loads the active defect set and per-profile guardrails for a run, and grows the living catalog.

Builds the story and information architecture so the deck is right *before* it is pretty. Prevents content-layer defects (L2) at the source. > **Content design comes first — empty space is a content failure, not a style > choice.** Before any layout, every planned element must be *earned* by enough > real substance to fill its space: a concrete fact, a number, a named entity, an > evidence detail.

Turns the narrative outline into a typed IR scene-graph and prevents geometric defects at generation time (prevention > detection).

The validation step: render the deck and measure it against the Defect DB. This is the trust core — deterministic L1/L2 gates, not opinion.

The correction step: fix only the defects in the report, then re-measure. Converges or stops; never rewrites the whole deck.