Treat the learning process as qualitative research with screenwriters, comic-drama studios, and directors jointly participating, rather than word-frequency mining. First understand how a complete project conveys dramatic intent into executable text production specifications, then decide which knowledge can be transferred.
Lock Down Boundaries Before Starting
Read non-public sources only when the maintainer actively invokes this skill and the current session has already provided authorized read-only access. Keep connection details, credentials, source names, and private cards in an isolated workspace designated by the maintainer; they must not be written into this skill, public skill trees, public candidates, or evaluation fixtures. If there is no authorization or read-only status cannot be confirmed, stop reading sources and use only synthetic materials to rehearse the process.
This process studies only text and its version relationships. It may read creative briefs, story structures, scripts, asset/continuity ledgers, image prompts, storyboard/keyframe text, video prompts, and review and revision records; it does not generate, render, download, or inspect images, audio, or video. The public runtime never connects to non-public sources, nor does it undertake any media generation.
Core Principles
- Let the agent judge genre context, character desires, causality, information gaps, emotional progression, performance actions, shot intent, and production trade-offs.
- Entrust only deterministic steps such as format validation, file comparison, duplicate checking, and privacy scanning to existing tools; do not write rule-matching scripts for creative semantics.
- Statistics are used only to describe optional samples, discover coverage gaps, find anomalies, and arrange the next reading; creative rules must not be inferred directly from frequency, ranking, or correlation.
- A piece of candidate knowledge must simultaneously carry evidence, counterexamples or failure cases, conflicting explanations, applicable boundaries, and confidence; "common" does not equal "effective."
- Propose only the smallest reviewable public change each time; without blind testing and independent review, it must not enter the public skill directly.
Workflow
1. Establish read-only authorization and an isolated workspace
Record the purpose of this round, the text scope allowed to be read, content prohibited from being read, output location, and destruction requirements. Test one read-only behavior, but do not echo connection information. Establish a private research directory and a separate de-identified candidate directory; the public directory remains unchanged at first.
2. Select samples based on coverage gaps
First inspect the minimal index or directory information permitted by the maintainer. Look for gaps, control groups, and anomalous samples by genre, audience promise, core mechanism, length stage, production form, visual language, scene function, prompt responsibility, degree of completion, and revision status. Statistics can only help "what to read next," not answer "how it should be written." Prioritize forming differentiated purposive samples, rather than reading only the most popular, most common, or most easily queryable projects.
3. Qualitatively read through the complete project chain
For each project, read from upstream intent to downstream text delivery: creative/adaptation constraints → story engine and episode design → single-episode script → characters, scenes, props, and continuity → image prompts → storyboard and keyframe text → video prompts → review, revision, and delivery notes. If a layer does not exist, explicitly record it as missing; do not fill it in with another project or common sense.
Track along the same promise: how expectation is established, how obstacles and information gaps are created, what performable actions the characters use to advance, how the turn is set up, and how the shots/prompts preserve rather than rewrite the dramatic intent. Write cards only after reading the full text; do not substitute batch keyword hits for reading. For card fields and a synthesis example, see cards-and-coverage.md。
4. Write private observation / decision cards
First write an observation card, strictly distinguishing direct observation, agent interpretation, and the unknown; then write a decision card, recording whether a candidate formed, why the scope of applicability was narrowed, and what alternative explanations exist. Each card uses a stable internal reference that does not leak content to bind to the same project chain, but private cards never enter the public skill package.
5. Build a genre × mechanism coverage matrix
Use "genre/audience promise" as one axis, and use mechanisms such as "expectation establishment, conflict escalation, information release, relationship turn, reversal payoff, episode-end traction, performance action, shot continuity, continuity" as the other axis; prompt research then layers on production form/visual language, scene function, prompt responsibility, and version role. Each cell records supporting cards, counterexample cards, boundaries, gaps, and the next sample, without filling in occurrence counts or success rates. For the matrix format, see cards-and-coverage.md。
6. Compress conclusions using counterexamples, conflicts, and boundaries of applicability
Actively look for failures of the same mechanism, exceptions in the same genre, alternative explanations for the same result, and cases that work upstream but are not producible downstream. When a conflict is unexplained, keep it as a candidate rather than taking a majority vote. Rewrite "always true" as "under these audience promises, character states, and production constraints, this mechanism may solve this problem; when these signals appear, switch methods or do not use it." For detailed criteria, see synthesis-and-promotion.md。
7. De-identify and de-copy
Delete or abstract source names, character names, proprietary settings, original sentences, rare combinations, precise numerical values, paths, addresses, identifiers, and connection information. Break surface-level plot devices into "audience state → dramatic problem → optional mechanism → observable effect → failure boundary," then rebuild the example with entirely new genres, character relationships, scenes, and wording. If it can only be expressed by restating the original content, it does not enter the public candidates.
8. Form a public reference / rubric / synthetic fixture proposal
Each candidate package simultaneously proposes: a concise reference change, a rubric that can locate evidence and failure modes, a fully synthetic fixture that is not close to any source, as well as the tasks expected to benefit and potential side effects. The proposal first stays in the maintainer staging area and does not directly modify the public suite. Public candidates must not require runtime access to non-public sources, and must not claim that the text can prove real media effects.
9. Do a fresh-agent blind forward eval
Have a fresh agent that has not read the private cards, the original project, or the expected answer complete the same synthetic task using only the public baseline or one of the anonymous candidates. Blind-evaluate its dramatic judgment, genre fit, production executability, variation ability, and whether it mechanically applies templates. Execute blind-forward-eval.md; do not leak the candidate author's conclusions into the prompt.
10. Hand off to an independent reviewer
A reviewer who did not participate in this round of extraction, synthesis, or candidate writing rechecks evidence sufficiency, counterexamples, boundaries of applicability, de-identification, de-copy, blind-test design, and public operating boundaries. Candidate authors must not self-review and approve; unresolved leakage, near-replication, mechanization, or evidence gaps mean it is sent back.
11. promotion / retire
Only when a candidate passes fresh-agent blind forward eval and independent reviewer, and the reference, rubric, and synthetic fixture are mutually consistent, may the maintainer promote the minimal change to public materials. Keep decision records and rollback notes that contain no private content. When later counterexamples prove the candidate is too broad, shows no gain in blind testing, or induces templating, narrow, replace, or retire it; do not conceal knowledge changes through silent deletion.
Read on demand
- Read when writing private cards, synthetic examples, or a coverage matrix cards-and-coverage.md。
- Read when judging whether de-identification is possible, how to handle conflicts, or preparing promotion / retire synthesis-and-promotion.md。
- Read when designing fresh-agent blind forward eval, blind review, and reviewer handoff blind-forward-eval.md。
- Read when preparing promotion / hold / retire, or when the evaluation and rollback evidence for public rules needs to be replayable promotion-ledger.md。
- Read when learning different subjects, art styles, scenes, and image/video prompt mechanisms, or when distinguishing revision requests, input reference observations, and generated result observations prompt-production-learning.md。
Do not preload all references; read only the one needed for the current stage.