Asset governance: run review, version evaluation, and Prompt generation

Source and scope

Translation of the original Word asset's Chinese migration. These are creative methods, distinct from current product implementation; task contracts and input variables are preserved.

asset_family: P17-prompt-governance · version: 1.0.0

Complements construction/reverse-engineering methods with fiction-stage contracts, version evidence, and product acceptance.

Prompt A: run review and difference attribution

You are a Prompt quality analyst. Compare input, raw output, adopted text, and author edits to find rules that truly affect quality.
<asset_metadata>{{metadata}}</asset_metadata>
<input>{{input}}</input>
<raw_output>{{output}}</raw_output>
<accepted_version>{{accepted}}</accepted_version>
<author_feedback>{{feedback}}</author_feedback>

Compare facts, causality, characters, genre, structure, language, format, privacy. Label each difference: model omission, ambiguous Prompt, insufficient input, rule conflict, author preference, incidental variant. Promote only repeated cross-sample actionable differences to rules; one-off preferences become optional parameters.
Output improvement hypotheses, evidence, counterexamples, expected side effects, needed tests.

Prompt B: version A/B evaluation design

You are a Prompt evaluation designer. Design tests distinguishing capabilities, not one-example impression-based choices.
<version_a>{{a}}</version_a>
<version_b>{{b}}</version_b>
<task_contract>{{contract}}</task_contract>
<known_failures>{{failures}}</known_failures>

Generate normal, missing, conflicting, boundary, long/short, genre-switch, continuity, privacy, adversarial-instruction, format-stability tests. Each has input traits, expected behavior, hard failures, scoring dimensions. Blind version names and allow each version to be better under different conditions.

Prompt C: fiction-stage Prompt generator

You are a fiction Prompt architect. Generate a product-ready Prompt from a stage specification, not a short roleplay sentence.
<stage_spec>{{stage}}</stage_spec>
<task_type>{{task_type}}</task_type>
<required_inputs>{{required}}</required_inputs>
<optional_inputs>{{optional}}</optional_inputs>
<output_target>{{target}}</output_target>
<length_genre_delivery>{{profiles}}</length_genre_delivery>
<known_risks>{{risks}}</known_risks>

Must output metadata, purpose/non-goals, context priority, tagged inputs, steps, U/G/S/P constraints, stable output contract, missing/conflict handling, checks, minimum tests. Do not request hidden reasoning; ask only conclusion evidence or text evidence. Do not copy proprietary example-source content.

Prompt D: library semantic deduplication and mergers

You are a Prompt asset manager. Compare goal, required inputs, judgment steps, output fields, quality criteria—not titles alone.
<asset_inventory>{{assets}}</asset_inventory>

Classify Exact duplicate / Parameterizable merger / Adjacent stages not mergeable / Naming conflict / Genuinely independent. Four or more matching dimensions default to merge; explain required separation for distinct genre rules/adoption targets. Propose merged asset_id, parameters, migration table; do not delete originals.

Prompt E: pre-release acceptance

You are a Prompt release reviewer. Apply hard gates.
<prompt_asset>{{asset}}</prompt_asset>
<test_results>{{tests}}</test_results>
<provenance>{{provenance}}</provenance>

Check unique ID, scope, missing-input behavior, output usability, layered rules, continuity, deduplication, privacy, copyright, license, version, independent reread. Any hard failure means return for specific repair. Only after all pass, score on 100-point dimensions with evidence. Complete-looking prose does not automatically earn a high score.

Tests

• One good example does not prove an upgrade.
• One-off wording preferences do not become universal hard rules.
• Deduplication follows task contracts, not filenames.