Four-skill presentation system: ingest talks into a rhetoric vault, run interactive clarification, generate a speaker profile, then create new presentations that match your documented patterns. Includes a 102-entry Presentation Patterns taxonomy (91 observable, 11 unobservable go-live items) for scoring, brainstorming, and go-live preparation.
97
94%
Does it follow best practices?
Impact
98%
1.24xAverage score across 30 eval scenarios
Advisory
Suggest reviewing before use
Rule of thumb: if the model must draw something new, regenerate. If it only erases existing content, edit.
Every image edit prompt MUST include:
"DO NOT add any new elements." — suppresses Gemini's aggressive decoration"Let background continue naturally — no parchment patch." — prevents flat-fill artifacts when erasing"Keep the [X]. Keep the [Y]." — prevents unintended removal of nearby elementsThe --edit and --fix commands auto-append #1 and #2, but #3 must always
be added manually.
Never simplify or rewrite the speaker's original style anchor when iterating. The specificity of the anchor (period vocabulary, document conventions, material constraints) is what produces a coherent style across slides; pruning it for "cleanliness" reverts the output to a generic illustrative default.
Append to the anchor for new constraints — don't replace.
Builds (slide-by-slide reveals where elements appear progressively) are the hardest illustration case because each step must visually match the others.
Chain backwards from the full image:
N (final) = full slide imageN-1 = full minus the last element to reveal00 = empty frame (title + borders only, no content)Each step's input is the PREVIOUS step's output (chained edits work because the per-step diff is small and the style is preserved on erasure — see the edit-vs-regenerate asymmetry rule above).
Don't:
Naming convention: builds/slide-NN-build-MM.jpg where MM is the stage
index (00 is empty, 01 adds the first element, etc.).
For checklists or progressive-state slides (e.g., a form filling in across multiple slides), use the same approach: take the most-complete version as the base and image-edit backwards to earlier states. Adding checkmarks via edit preserves visual consistency far better than regenerating from prompt.
When an image is "almost perfect" — say, one mislabeled callout or a single extra element — do a targeted image-edit fix pass. Regenerating from scratch risks losing all the things that already work. The edit-only-removes asymmetry applies: erase the wrong element, then if needed regenerate just the corrected element.
Save experiment versions (v2, v3, v4) instead of overwriting the
working file. Slide illustrations take many iterations to converge; stomping
on a near-good output to try one more variation loses information that may
have been worth keeping. The disk cost of saving all attempts is trivial
compared to the cost of regenerating a known-good output you accidentally
clobbered.
evals
scenario-1
scenario-2
scenario-3
scenario-4
scenario-5
scenario-6
scenario-7
scenario-8
scenario-9
scenario-10
scenario-11
scenario-12
scenario-13
scenario-14
scenario-15
scenario-16
scenario-17
scenario-18
scenario-19
scenario-20
scenario-21
scenario-22
scenario-23
scenario-24
scenario-25
scenario-26
scenario-27
scenario-28
scenario-29
scenario-30
rules
skills
presentation-creator
references
patterns
build
deliver
prepare
scripts
vault-clarification
vault-ingress
vault-profile