Design a high-quality PersonWise course
Build the blueprint before creation
Write a concise, secret-free blueprint:
- learner and prerequisite knowledge;
- observable learner outcome;
- course class:
concrete-object,open-ended, or an explicitly partitioned hybrid; - dominant teaching archetype from
course-archetypes.md; - factual authority: open research, strict supplied materials, or source-assisted research;
- language and earned page count;
- page-by-page teaching arc;
- examples, comparisons, diagrams, tables, illustrations, and guided practice that earn their space;
- visual system and prohibited motifs;
- spoken style and Voice-delivery brief;
- presenter brief covering the teaching job, warmth/authority, energy, formality, framing, and language compatibility;
- final target: draft, private, link, or separately authorized Topics review.
Choose page count from the learner promise:
- 4–8 pages for one bounded workflow or interpretive lens;
- 8–16 for a focused lesson;
- 16–26 for mechanisms, examples, counterarguments, practice, and synthesis;
- 26–30 only when the depth is real.
Do not pad. Each page needs one primary teaching job and a reason for its visual form.
Understand the creation fields
| Field | Public meaning |
|---|---|
name |
Visible course title, at most 200 characters. |
topic |
Compact course constitution, at most 200 characters. It strongly guides Outline, metadata, style, and—when content is absent—Narration. |
description |
Learner-facing summary, at most 500 characters. |
target_audience |
Audience, prerequisites, context, and success criterion, at most 500 characters. |
content |
Aligned long-form generation input, at most 8000 characters. It does not replace uploaded strict sources. |
language |
Course language/locale; also constrains presenter Voice compatibility. |
desired_slide_count |
Fixed page count from 1–30. Count/order cannot be changed through this public workflow after creation. |
style_instruction |
Global visual system, at most 2000 characters: palette, medium, typography character, composition rhythm, and constraints. |
script_style |
Spoken teaching behavior, at most 2000 characters: tone, pacing, explanation structure, analogy policy, and narration exclusions. |
voice_render_style_instruction |
TTS delivery direction, at most 300 characters; no factual content. |
avatar_id |
Optional known compatible presenter UUID. Prefer later catalog review rather than guessing. |
voice_id / voice_plane_voice_id |
Optional known compatible Voice lane UUID. Never invent or mix an unrelated presenter/Voice pair. |
knowledge_source_mode |
open or materials_only; fixes the factual acquisition boundary. |
declared_sources |
Exact retained document count expected before deferred project creation. |
visual_review_capability |
Declare multimodal only when this Agent can truly inspect downloaded review images; otherwise none. |
distribution_target |
draft, private, link, or topics_review, bounded by OAuth consent and advertised capabilities. Always set it explicitly: when omitted it resolves to the OAuth grant's publication ceiling, which can be link; use private unless the user asked for broader access. |
At least one of topic, content, or a declared source is required. materials_only requires at
least one declared source.
Respect topic and content precedence
The generation pipeline uses topic before content for the Outline and project initializers, but
uses content before topic for Page text and Narration. Therefore:
- Use
topicalone for one compact instruction that must propagate consistently. - When long
contentis needed, make it semantically identical to thetopicconstitution. - Put audience, visual style, script behavior, and Voice delivery in their dedicated fields.
- Never place conflicting instructions in
topicandcontent. - Uploaded strict materials govern facts;
topic/contentstill govern selection, organization, teaching intent, and exclusions.
A useful compact topic contains:
subject + learner promise + arc + required visual grammar + exclusions
Example:
Forces that shape Earth's surface (internal forces → external forces → landform evidence →
human consequences; use cross-sections, process diagrams, and comparisons; one causal relationship
per page; avoid tourism-poster scenery)
Choose the knowledge boundary
Open research
Use knowledge_source_mode=open when the model may select and synthesize generally available
knowledge. Sources may be absent.
Review concrete claims for accuracy and currency. Open mode is not permission to invent exact metrics, quotations, product interfaces, chronology, or attribution.
Strict supplied materials
Use materials_only when correctness means staying within the supplied PDF, PPTX, DOCX, Markdown,
or TXT files.
Before creation:
-
Retain only authoritative files.
-
Set
declared_sourcesto their exact count. -
Build a source map:
source -> authoritative sections/claims -> intended pages -> claims that must not be added -
Request and complete one document upload ticket per source.
-
Wait until every retained source reaches canonical complete state.
Do not silently switch to open mode when a source fails. Strict mode is a generation constraint, not a compliance certificate: inspect every claim, remove unsupported additions, distinguish source statements from implications, and surface source disagreement instead of merging it into false certainty.
Source-assisted open research
Use open with documents when supplied facts should anchor the lesson but relevant outside context
is allowed. Label this honestly; do not describe it as strict source compliance.
Apply the editorial-authority boundary
Read course-archetypes.md before correcting model-authored content.
For a concrete-object course, verify the external object's facts, claims, chronology, UI,
capabilities, relationships, outputs, metrics, and source imagery. Plausible specificity is still
unsupported until evidence establishes it.
For an open-ended course, the model legitimately chooses structure, examples, analogies, and
interpretive emphasis within the brief. Preserve coherent authorship. Correct only:
- demonstrable factual or source error;
- copyright or attribution error;
- internal contradiction;
- clear pedagogical failure;
- approved-brief violation;
- material safety problem.
For a hybrid, mark which pages or claims use each authority boundary. An original film-analysis arc may be valid while release facts, credits, quotations, and scene details remain externally verifiable.
Before editing, record the objective category and evidence. “I prefer another example” is not a correction reason.
Design and review the Outline
Every page should:
- perform one teaching job;
- build on the previous page or create a deliberate contrast;
- contribute a distinct concept, example, practice, counterargument, or synthesis;
- contain enough factual support for its title and Key points;
- offer a meaningful visual job.
Review the full Outline, including every generated title. Hunt for:
- repeated claims in different words;
- a sequence that requires unexplained prerequisites;
- decorative pages with no learning function;
- a conclusion that introduces new core content;
- unsupported specificity or false certainty;
- plot summary, feature catalogues, or trivia replacing explanation;
- a page title that promises more than its Key points deliver.
Keep slide IDs stable. Do not add, delete, or reorder pages. Batch the smallest sufficient title and
key_points corrections, then fetch a fresh revision.
Align Title, Key points, Page text, and Narration
Treat the four fields as an aligned contract:
Title
Name the page's teaching job. Prefer a claim, question, mechanism, decision, or contrast over a generic section label.
Key points
Define what the page must teach. Use distinct, supportable claims. Key points are not a pile of keywords and should not duplicate each other.
Page text
Define what the learner sees and what primarily controls the generated slide image. Keep it concise, hierarchical, and layout-ready. State diagram labels, comparison dimensions, table headings, and critical caveats explicitly. Avoid transcript paragraphs.
Narration script
Explain the page: context, causal links, transitions, nuance, evidence limits, and examples. Do not merely read visible text. Keep the spoken claim set aligned with the page and source boundary.
Change-impact matrix
| Changed field | Recheck |
|---|---|
| Title | Key points, Page text, Narration, visual job |
| Key points | Title, Page text, Narration, learner-outcome alignment |
| Page text | Key points, Narration, generated image |
| Narration | Key points, Page text, factual/source alignment |
When one conceptual correction affects several fields, update them in one atomic revision-checked batch. After every accepted batch, fetch a fresh snapshot before the next mutation.
Design useful visuals
Specify the information job, not only a style adjective:
- process -> numbered flow, state diagram, or causal chain;
- mechanism -> labeled schematic or cross-section;
- comparison -> consistent columns with meaningful dimensions;
- chronology -> scaled or explicitly non-scaled timeline;
- hierarchy -> nested structure or dependency tree;
- evidence -> source-grounded annotated artifact or claim/evidence table;
- interpretation -> juxtaposed details, motifs, or competing readings;
- practice -> scenario, decision point, and feedback path.
Treat visual semantics as claims. A green check implies success; a rising chart implies growth; a padlock implies security; a polished dashboard implies an observed interface or dataset. Ban these when evidence does not support them. Avoid invented percentages, counts, scores, durations, customers, logos, maps, quotations, and pseudo-screenshots.
Use Reference images for subject, palette, composition, or visual grammar. Use Pin only when the exact supplied image should remain the hero visual. A Pin provider contract is not proof of pixel-perfect preservation.
Define publish quality
Before finishing, require:
- full Outline and Page text/Narration review;
- every retained strict source complete and mapped;
- no unresolved factual, source, safety, or internal-consistency finding;
- all slide images canonically complete;
- visual review recorded honestly as
completed,partial, ornot_performed; - one exact language-compatible presenter/Voice pair;
- requested layout persisted;
- successful first publish when authorized;
- link visibility only when requested/resolved;
playable=truebefore reporting public or embed URLs;- private delivery gives the login-required
editor_urland never calls it a share link; - Topics state described as a pending/reviewed submission, never as automatic distribution.