Angular Noir

Turn a familiar subject into a stark, hard-edged crime-strip illustration. A supplied face stays recognizable, only carved into confident planes; a described scene stays yours; a context becomes a case with one clear clue. Black does most of the talking. Deliver the image itself.

Required capabilities and resources

• Use an image-generation/editing tool. For a photograph, use an actual image edit with the photograph attached as the edit target. Do not reconstruct it from a description alone.
• Use netsi-prompt-generator for the internal image prompt. If it is installed, read its instructions. Otherwise read the bundled, complete copy at references/prompt-generator.md and apply that skill's workflow to the prompt subtask.
• Read references/style-profile.md before composing the prompt, and references/tweaks.md when adjustments or a comparison are relevant.
• Inspect the bundled generated examples in https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/ as anonymous visual calibration, especially https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/04-streetlamp-detective.png for black shapes and architecture and https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/01-pool.png for how a real face is carved into planes without losing identity. When the image tool supports multiple references, supply the input photograph as the edit target and a relevant bundled illustration as the style reference only. In a new scene, the bundled illustrations are style references only. Never copy their subject, composition, identity or labels into the user's result. Do not redistribute historical research scans.
• The prompt generator's output contract applies internally: its English prompt goes to the image tool. The parent skill's user-facing deliverable remains the finished image.
• If image generation is unavailable, say that clearly in the user's language. Do not label an English prompt or a mockup as a generated image. Never silently substitute a different provider or paid API.

Recognize one of three inputs

A — Image. Inspect the photograph. Record the person or subject, face proportions, age, expression, pose, hands, clothing, glasses, recognizable accessories, objects, framing and meaningful setting. Preserve these unless the user explicitly asks to change them. Retain the input aspect ratio by default. A style reference is a different role from the edit target; do not swap them.

B — Scene. Extract who or what is present, actions, setting, composition and any explicit constraints. Keep the scene's content. A scene that is not a crime scene keeps its subject: the noir vocabulary then appears through hard light, black shadow shapes, angular architecture and staging, not by adding guns, victims or police. If no aspect ratio is supplied, use 16:9 for a scene; follow a stated medium such as a square social post instead when provided.

C — Context. Extract the message, audience, mood and destination from the supplied text or conversation. Treat it as a small case: choose a concrete, original visual metaphor built around a detective-like search, a trail or one decisive clue. Do not invent facts, numbers, quotations or branding. Make the message understandable without text by default. Ask only if materially different interpretations prevent a useful result. Use the requested destination's aspect ratio, otherwise 16:9.

If inputs are combined, infer their roles: for example, an image is the subject and a paragraph supplies the illustration's message. Ask one focused question only when the roles conflict.

Optional adjustments: zero to three

The canonical controls are Angle, Color, and Clues. Values and defaults live in the tweaks reference. Never weaken the identity of this style by changing its heavy contour and solid blacks to a generic filter or another medium.

If the request is clear, apply the explicit values or defaults and generate immediately. Do not require a questionnaire. If the user asks to explore, requests a materially ambiguous adjustment, or a meaningful unresolved choice would change the result, use an AskUserQuestion interaction: the environment's interactive question tool, or the same concise question in chat when no such tool is available.

Offer at most three relevant tweak controls. Ask one focused question at a time, explain its visual effect in plain language, place the recommended/default answer first, and allow a custom answer and discussion. Where tool schemas add free text automatically, do not duplicate it with a fake option. Infer answers already supplied. Do not mistake silence for consent to a material change. If the user simply wants the default image, no question is needed.

Translate questions, explanations and choices into the user's language. Canonical parameter values stay English in prompts and grid labels. Natural-language requests such as "kun sort og hvid" may be mapped directly to a suitable value (Color = Ink) without asking again.

Respect a scene's stated lighting and content: a sunny beach scene keeps daylight and expresses noir through hard shadow shapes rather than turning to night, and a request for "no arrows" implies Clues = None. Stated content takes precedence over a style default.

Generate, inspect, deliver

1. Inspect the input and identify its type and invariants.
2. Load the detailed style profile. Resolve any relevant optional adjustments.
3. Apply netsi-prompt-generator to form one self-contained internal prompt with role, task, context/constraints and output format. Explicitly set the entire image-tool prompt to English, regardless of the dialogue language. Preserve requested in-image text verbatim in its requested language; this is the only language exception inside the English instructions.
4. Include the profile's core traits, their priorities, selected tweak values and the input's exact preservation constraints. If Clues is not None, choose the clue objects yourself from the subject or message and describe where the pointers go. Use descriptive traits only. Do not mention artist names, named fictional characters, existing works, signatures or imitation slogans.
5. Generate or edit the actual image. For a local edit target, inspect it before editing and pass the file through the image tool's supported reference mechanism.
6. Inspect the result: recognize the subject; check face, expression, accessories and hand gesture; check scene content; check the heavy unwavering contour, faceted planes, solid black shadow shapes, angular architecture, limited color and any clue pointers; check any requested lettering. The style must reshape the drawing, not merely raise contrast on a photograph.
7. If a material defect exists, make a focused revision and inspect again. Keep a consistent target identity and do not silently alter the scene to make the drawing easier.
8. Deliver the finished image, with at most a short useful caption in the user's language. Keep the internal prompt hidden unless the user explicitly asks for it. Save final project-bound assets in that project's folder; do not overwrite the source photograph.

Never turn a supplied person into a villain, a grotesque or a known comic character. Angular planes sharpen a real face; they do not distort its proportions, age or ethnicity, and they never mock. Grotesque exaggeration is reserved for invented background figures, and only when the user asks for it.

Comparison grids

When requested, create one 16:9 image with 4 columns × 3 rows (12 image cells). Use the same subject, identity, pose, objects and crop in every cell. Reserve a legible label strip under each image, inside the overall canvas. Labels contain the exact English parameter names and values, never artist names. Use the schedule in references/tweaks.md: it includes the default and varies Angle and Color while keeping Clues at None. A matrix may explore Clues instead if requested, but must state the exact changed values.

Generate the artwork with the image tool. Use supported layout/text composition tools to assemble a precise grid if available; do not use code to imitate the illustration or threshold a photograph as a substitute. A single image-tool call may render the complete labeled grid. Verify 12 cells, order, visible differences, literal labels and a 16:9 canvas (allow one pixel of integer raster rounding) before delivering. Regenerate if lettering or structure is wrong.

Do not claim pixel-identical identity or perfectly isolated experimental control: image generation can vary details. Aim for faithful, visually comparable variations and disclose a material limitation if it remains.

Start with an example

• Attach a photograph: "Apply this style. Keep the face, pose and surroundings recognizable."
• Describe a scene: "A night-shift detective in a long coat crouches to study a muddy footprint under a single streetlamp, while rain slants across a row of angular brick buildings."
• Supply context: "Illustrate tracking down the root cause of a production bug, for a team retrospective slide. No text in the image."
• Explore adjustments: "Show my attached lakeside portrait in a 4 × 3 comparison grid with English tweak labels."

The bundled examples are https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/01-pool.png and https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/02-car.png (photographs), https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/03-tweak-grid.png (the comparison grid), https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/04-streetlamp-detective.png (scene) and https://assets.netsi.dk/skills/netsi-style-like-angular-noir/images/05-root-cause-context.png (context). See EXAMPLES.md, when present, for source photos and settings.