Commit 1: editorial image prompt + style-template setting (#146)

* Commit 1: editorial image prompt + style-template setting

Pivot the image pipeline from the single-panel comic style to a text-free
editorial scientific illustration, and replace the Style Suffix setting with
an editable Style Template setting. Commit 1 of the in-plugin editorial image
pipeline (no composer yet; pipeline otherwise unchanged).

- New constant PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE (brief §7 editorial
  template with the single {{ topic_summary }} token).
- New setting prautoblogger_image_style_template (textarea) replaces the
  Style Suffix field; single-token validation on save (brief A5); old
  prautoblogger_image_style_suffix mirrored to _deprecated for one cycle and
  not deleted (one-time migration prautoblogger_migrated_style_template_v0160).
- REWRITER_SYSTEM_PROMPT now emits a 1-2 sentence editorial topic/mechanism
  summary as the SCENE (no comic gag), keeping the CAPTION line for the
  unchanged HTML-caption-below-image path; scene/caption parsing contract
  preserved.
- New PRAutoBlogger_Image_Template_Filler substitutes the scene into the
  template's {{ topic_summary }} slot across all three entry points; strips
  control chars + clamps length (A5); degrades safely on missing token or
  empty summary so a blank/token-only prompt is never sent to the provider
  (A5/A6).
- Version bump 0.15.1 -> 0.16.0 + CHANGELOG + ARCHITECTURE.md updates.

Agent-Session: image-editorial-commit1

* style: auto-fix PHPCS violations via phpcbf

---------

Co-authored-by: peptiderepo <peptiderepo@users.noreply.github.com>
This commit is contained in:
peptiderepo
2026-06-09 13:18:38 +07:00
committed by GitHub
parent a4f3b1832e
commit d00e746727
11 changed files with 335 additions and 60 deletions

View File

@@ -6,13 +6,16 @@ declare(strict_types=1);
*
* Builds image prompts from article content or source data.
*
* Uses a cheap LLM (via OpenRouter) to distill article content into concise
* visual scene descriptions optimised for image generation. Falls back to rule-based
* Uses a cheap LLM (via OpenRouter) to distill article content into a concise
* 1-2 sentence editorial topic/mechanism summary, then substitutes that summary
* into the admin-editable editorial style template (the {{ topic_summary }}
* slot) to produce a text-free image-gen prompt. Falls back to rule-based
* synthesis if the LLM call fails, so image generation never blocks on a
* prompt-rewriting outage.
*
* Triggered by: PRAutoBlogger_Image_Pipeline during content generation run.
* Dependencies: PRAutoBlogger_OpenRouter_Provider (LLM call),
* PRAutoBlogger_Image_Template_Filler (template substitution),
* PRAutoBlogger_Cost_Tracker (logs prompt-rewrite cost),
* PRAutoBlogger_Logger (diagnostics).
*
@@ -27,34 +30,39 @@ class PRAutoBlogger_Image_Prompt_Builder {
* prompts. Exposed as a `public const` so the admin-settings layer can
* use it as the default value for `prautoblogger_image_prompt_instructions`.
* The option, when non-empty, wins at call time (see rewrite_via_llm).
*
* Editorial pivot (v0.16.0): the model emits a concise topic/mechanism
* summary as the SCENE — concrete, single centered focal subject, no text,
* no people-as-gag, no logos — which the prompt builder substitutes into
* the editorial style template's {{ topic_summary }} slot. The CAPTION line
* is preserved unchanged for the existing HTML-caption-below-image path.
*/
public const REWRITER_SYSTEM_PROMPT = <<<'PROMPT'
You are a comedy writer and single-panel cartoon creator, like Gary Larson (The Far Side) meets science humor.
You write subject descriptions for a text-free editorial science illustration about peptides, supplements, or biohacking.
Given an article title and summary about peptides, supplements, or biohacking, create a SINGLE-PANEL COMIC concept. Output TWO parts separated by a blank line:
Given an article title and summary, output TWO parts separated by a blank line:
SCENE: One sentence describing the visual gag — a funny, absurd, or ironic situation related to the article's topic. Favor CONCRETE, HUMAN-SCALE scenes: people reacting to things, doctors and patients, scientists in a lab, gym bros with peptide vials, partners at a kitchen table, ordinary objects in ordinary rooms. Keep it simple — one clear visual joke, 1-3 characters max.
SCENE: A concise 1-2 sentence description of the article's core topic or mechanism, written as the SUBJECT of a single editorial illustration. Name ONE concrete, centered focal subject and the few supporting visual elements around it (e.g. a labelled-free vial and dropper, a stylised receptor on a cell membrane, a molecular chain, a dosing syringe, an organ cross-section). Favor a clear central focal point with margin around it so the image crops well. This is image-gen direction, not prose.
CAPTION: A short, punchy caption or speech bubble line (under 15 words) that delivers the punchline. The humor should be dry, nerdy, and accessible — the reader should chuckle even if they only half-understand the science.
CAPTION: A short, informative caption line (under 15 words) for display as HTML text below the image. Plain and editorial, not a joke.
Rules:
- The joke MUST relate to the article's actual subject matter, not generic science humor
- DO NOT personify molecules, peptides, hormones, proteins, or other chemical structures (no "anthropomorphic molecule", no "cartoon hormone"). Image models render anthropomorphized chemistry as incoherent blobs. If the article is about a substance, show a PERSON reacting to or using the substance; if the article is abstract (a mechanism, a process, a concept), fall back on a human observer reacting to its effect.
- Characters can have simple cartoon faces (round heads, dot eyes, expressive eyebrows)
- Keep the scene physically simple — few objects, clear staging, easy to read at thumbnail size
- The caption should work as a standalone joke even without the image
- No logos, watermarks, or branding
- The SCENE must depict the article's actual subject/mechanism, not a generic science scene.
- NO text, words, captions, speech bubbles, labels, or logos described in the SCENE — the illustration is text-free; text is rendered separately.
- NO people used as a visual gag and NO cartoon/comic framing. A human anatomical element (a cell, an organ, a hand holding a vial) is fine if it is the legitimate subject.
- Do NOT personify molecules, peptides, hormones, or proteins (no "smiling molecule"); depict them as clean stylised diagrams or structures.
- Keep the staging simple with one clear focal subject, readable at small sizes.
- Output ONLY the scene and caption. No preamble, no explanation.
Example output format:
A muscular gym bro stares, deflated, at a tiny glass vial in his hand while his friend squats a barbell effortlessly in the background.
A single glass peptide vial on a clean surface with a fine dropper above it, a faint stylised molecular chain arcing behind as a backdrop, centered with generous negative space.
"They told me the peptide did the work, not the squats."
A concentrated look at how the compound is prepared before use.
PROMPT;
/**
* Max tokens for the rewriter response. Comic concepts need more room
* for the scene description plus the caption punchline.
* Max tokens for the rewriter response. Enough room for the 1-2 sentence
* topic/mechanism summary plus the short editorial caption line.
*/
private const REWRITER_MAX_TOKENS = 180;
@@ -85,8 +93,9 @@ PROMPT;
* Build a visual prompt from finished article content.
*
* Tries LLM rewriting first; falls back to rule-based synthesis on
* failure. Splits scene (for image gen) from caption (HTML below the
* image) and appends the style suffix.
* failure. Splits scene (the topic summary, for image gen) from caption
* (HTML below the image) and substitutes the scene into the editorial
* style template's {{ topic_summary }} slot.
*
* @param array{post_title?: string, post_content?: string, suggested_title?: string} $article_data
* @return array{prompt: string, caption: string}
@@ -98,10 +107,8 @@ PROMPT;
$parsed = $this->rewrite_via_llm( $title, $first_para );
$style_suffix = $this->get_style_suffix();
return array(
'prompt' => trim( $parsed['scene'] . ' ' . $style_suffix ),
'prompt' => PRAutoBlogger_Image_Template_Filler::fill( $parsed['scene'] ),
'caption' => $parsed['caption'],
);
}
@@ -120,10 +127,8 @@ PROMPT;
$parsed = $this->rewrite_via_llm( $title, $context );
$style_suffix = $this->get_style_suffix();
return array(
'prompt' => trim( $parsed['scene'] . ' ' . $style_suffix ),
'prompt' => PRAutoBlogger_Image_Template_Filler::fill( $parsed['scene'] ),
'caption' => $parsed['caption'],
);
}
@@ -260,10 +265,9 @@ PROMPT;
* @return array{prompt: string, caption: string}
*/
public function build_fallback_prompt( string $title ): array {
$parsed = PRAutoBlogger_Image_Scene_Parser::synthesize_visual_concepts_fallback( $title, '' );
$style_suffix = $this->get_style_suffix();
$parsed = PRAutoBlogger_Image_Scene_Parser::synthesize_visual_concepts_fallback( $title, '' );
return array(
'prompt' => trim( $parsed['scene'] . ' ' . $style_suffix ),
'prompt' => PRAutoBlogger_Image_Template_Filler::fill( $parsed['scene'] ),
'caption' => $parsed['caption'],
);
}
@@ -281,20 +285,4 @@ PROMPT;
$override = (string) get_option( 'prautoblogger_image_prompt_instructions', '' );
return '' !== trim( $override ) ? $override : self::REWRITER_SYSTEM_PROMPT;
}
/**
* Get image style suffix with empty-string safeguard.
*
* Treats both absent and empty string as "use default" to prevent
* silent style loss if the admin settings field is saved empty.
*
* @return string The style suffix with safe fallback.
*/
private function get_style_suffix(): string {
$style_suffix = (string) get_option( 'prautoblogger_image_style_suffix', '' );
if ( '' === trim( $style_suffix ) ) {
$style_suffix = PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_SUFFIX;
}
return $style_suffix;
}
}

View File

@@ -70,22 +70,25 @@ class PRAutoBlogger_Image_Scene_Parser {
/**
* Rule-based fallback when LLM rewriting is unavailable.
*
* Kept deliberately simple — this only runs if OpenRouter is down.
* Kept deliberately simple — this only runs if OpenRouter is down. Produces
* a text-free editorial-illustration subject (not a comic) so the prompt
* builder always has a non-empty topic summary to substitute into the
* style template, never shipping a blank prompt to the provider (brief A6).
*
* @param string $title Main heading / topic.
* @param string $supporting_text Additional context.
* @return array{scene: string, caption: string} Scene for image gen, caption for HTML.
*/
public static function synthesize_visual_concepts_fallback( string $title, string $supporting_text ): array {
// Fallback produces a simple comic concept when the LLM is unavailable.
// Fallback produces a simple editorial-illustration subject when the LLM is unavailable.
$scene = sprintf(
'A cartoon scientist in a lab coat looking bewildered while examining something related to: %s.',
'A clean, centered editorial science illustration representing the topic of %s, with a single clear focal subject and generous negative space.',
$title
);
return array(
'scene' => trim( $scene ),
'caption' => 'Science is full of surprises.',
'caption' => '',
);
}
}

View File

@@ -0,0 +1,185 @@
<?php
declare(strict_types=1);
/**
* phpcs:ignore WordPress.Files.FileName.InvalidClassFileName -- class naming convention differs from WordPress standard
*
* Fills the admin-editable editorial image-prompt template with the
* rewriter-produced topic summary, replacing the old "scene + style suffix"
* concatenation.
*
* The template (option `prautoblogger_image_style_template`, default
* PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE) contains exactly one
* `{{ topic_summary }}` token. This class:
* - resolves the active template (admin override wins; blank → default),
* - sanitises and length-clamps the topic summary (brief A5),
* - substitutes the token, and
* - degrades safely (brief A5/A6) so a blank or token-only prompt is never
* emitted to the image provider.
*
* Who calls it: PRAutoBlogger_Image_Prompt_Builder (static call from each of
* build_article_prompt / build_source_prompt / build_fallback_prompt).
* Dependencies: PRAutoBlogger_Logger (diagnostics only).
*
* @see core/class-image-prompt-builder.php — Sole caller.
* @see ARCHITECTURE.md — Image generation data flow.
*/
class PRAutoBlogger_Image_Template_Filler {
/** The single substitution token the template must contain. */
public const TOKEN = '{{ topic_summary }}';
/**
* Max characters retained from the rewriter topic summary before it is
* substituted into the template. Keeps the final provider prompt bounded
* even if the rewriter ignores its length guidance.
*/
public const MAX_SUMMARY_CHARS = 320;
/**
* Build the final image-generation prompt from a rewriter scene.
*
* @param string $scene Raw topic/mechanism summary from the rewriter (the
* "scene" half of the scene/caption contract).
* @return string Provider-ready prompt; never empty.
*/
public static function fill( string $scene ): string {
$template = self::resolve_template();
$summary = self::sanitize_summary( $scene );
// Brief A6: never ship a blank or token-only prompt. If the rewriter
// produced nothing usable, return the template with the token removed
// only as a last resort — but the caller guards this by passing a
// rule-based fallback scene, so an empty summary here is unexpected.
if ( '' === $summary ) {
PRAutoBlogger_Logger::instance()->warning(
'Editorial prompt fill received an empty topic summary; emitting style-only prompt.',
'image_template_filler'
);
return self::strip_token( $template );
}
$token_count = substr_count( $template, self::TOKEN );
if ( 1 === $token_count ) {
return trim( str_replace( self::TOKEN, $summary, $template ) );
}
// Brief A5: admin removed or duplicated the token. Fall back to
// appending the summary to the style text and log a warning rather
// than shipping a malformed prompt.
PRAutoBlogger_Logger::instance()->warning(
sprintf(
'Image style template has %d "%s" tokens (expected 1); appending summary instead of substituting.',
$token_count,
self::TOKEN
),
'image_template_filler'
);
$base = self::strip_token( $template );
return trim( $base . ' ' . $summary );
}
/**
* Sanitise and validate the admin-submitted style template on save.
*
* Multi-line, so sanitised with sanitize_textarea_field() (not
* sanitize_text_field(), which would strip newlines). Per brief A5 the
* template must contain exactly one TOKEN; otherwise the previous value is
* kept and a settings error is surfaced. A blank submission falls back to
* the editorial default so image generation can never be bricked.
*
* @param string $value Submitted template.
* @return string Sanitised, validated template.
*/
public static function sanitize_for_save( string $value ): string {
$clean = sanitize_textarea_field( $value );
if ( '' === trim( $clean ) ) {
return PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE;
}
$token_count = substr_count( $clean, self::TOKEN );
if ( 1 !== $token_count ) {
add_settings_error(
'prautoblogger_image_style_template',
'prautoblogger_image_style_template_token',
sprintf(
/* translators: 1: required token, 2: number of tokens found. */
esc_html__( 'The Style Template must contain exactly one %1$s token (found %2$d). Keeping the previous value.', 'prautoblogger' ),
esc_html( self::TOKEN ),
(int) $token_count
)
);
return (string) get_option( 'prautoblogger_image_style_template', PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE );
}
return $clean;
}
/**
* Resolve the active template: admin override wins; blank → default.
*
* @return string Non-empty template string.
*/
private static function resolve_template(): string {
$override = (string) get_option( 'prautoblogger_image_style_template', '' );
if ( '' !== trim( $override ) ) {
return $override;
}
return PRAUTOBLOGGER_DEFAULT_IMAGE_STYLE_TEMPLATE;
}
/**
* Strip control characters and clamp length of the rewriter summary
* (brief A5). Whitespace runs (including newlines) are collapsed to single
* spaces so the summary sits cleanly inside the template.
*
* @param string $scene Raw rewriter scene.
* @return string Sanitised, length-clamped summary (possibly empty).
*/
private static function sanitize_summary( string $scene ): string {
// Remove control characters (C0/C1) except via whitespace collapse below.
$clean = preg_replace( '/[\x00-\x1F\x7F]+/u', ' ', $scene );
if ( null === $clean ) {
// preg_replace can return null on bad UTF-8; degrade conservatively.
$clean = preg_replace( '/[\x00-\x1F\x7F]+/', ' ', $scene );
$clean = is_string( $clean ) ? $clean : '';
}
// Collapse whitespace runs to single spaces.
$collapsed = preg_replace( '/\s+/u', ' ', $clean );
$clean = is_string( $collapsed ) ? $collapsed : $clean;
$clean = trim( $clean );
if ( '' === $clean ) {
return '';
}
// Clamp length (multibyte-aware where available).
if ( function_exists( 'mb_strlen' ) ) {
if ( mb_strlen( $clean ) > self::MAX_SUMMARY_CHARS ) {
$clean = rtrim( mb_substr( $clean, 0, self::MAX_SUMMARY_CHARS ) ) . '...';
}
} elseif ( strlen( $clean ) > self::MAX_SUMMARY_CHARS ) {
$clean = rtrim( substr( $clean, 0, self::MAX_SUMMARY_CHARS ) ) . '...';
}
return $clean;
}
/**
* Remove every occurrence of the token from a template, tidying the
* leftover whitespace so the residual style text is still well-formed.
*
* @param string $template Template possibly containing the token.
* @return string Token-free, trimmed template.
*/
private static function strip_token( string $template ): string {
$stripped = str_replace( self::TOKEN, '', $template );
$stripped = preg_replace( '/[ \t]+/', ' ', $stripped );
$stripped = is_string( $stripped ) ? $stripped : $template;
return trim( $stripped );
}
}