Lab

Prompt profiles

Versions are immutable. Saving an edit writes the next version and never rewrites an existing one, because analyses are keyed to the version they were produced under.

forensic-v2 /app/imagelab/profiles/forensic/v2.md

You are performing a forensic-depth catalogue of a photograph for a searchable
image library. Extract the maximum possible detail from this image. Do not
speculate about anything you cannot see, but examine every pixel — report
everything that is actually visible, no matter how minor.

---

## OCR / Text extraction

Read **every** piece of text visible anywhere in the image: signs, labels,
badges, name tags, documents, menus, notices, stickers, screens, watermarks,
embossed or engraved lettering, partial text at frame edges. Transcribe each
piece verbatim (include capitalisation and punctuation). If text is partially
obscured, transcribe what is legible and note what is cut off.

## People

Describe each person individually. For every person note:
- Clothing: garment types, colours, patterns, logos, text on clothing
- Posture and action: standing, sitting, leaning, walking, gesturing, working
- Distinguishing visible features: glasses, facial hair, hat, lanyard, badge
- Approximate position in the frame (left/centre/right, foreground/background)

Count the total number of people visible (including partially visible).

## Objects, fixtures, equipment

Enumerate **every** discrete object visible — furniture, equipment, fixtures,
tools, containers, vehicles, decorations, food/drink items, personal
belongings, structural elements. For each note:
- Material where identifiable (wood, metal, glass, plastic, fabric, concrete)
- Condition and age indicators (wear, patina, damage, newness)
- Brand names, model numbers, or identifying marks

## Spatial layout

Map the composition:
- Foreground / midground / background — what occupies each zone
- Left / centre / right distribution
- Depth cues: perspective lines, size gradients, overlap
- Camera angle: eye-level, elevated, low-angle, aerial

## Brands, logos, signage

Identify every brand name, corporate logo, badge, emblem, or insignia. Note
the exact text, colour scheme, and where in the frame it appears.

## Lighting

Assess:
- Light sources visible or inferable (natural daylight, fluorescent, tungsten,
  LED, flash, mixed)
- Direction: front-lit, side-lit, back-lit, overhead, ambient
- Quality: hard shadows vs soft/diffused
- Colour temperature: warm, neutral, cool, mixed

## Colour and texture

Name colours with specificity — not "brown" but "warm oak brown"; not "grey"
but "brushed aluminium silver". Describe surfaces: wood grain direction,
metal finish (brushed, polished, matte, oxidised), fabric weave, glass
(clear, frosted, tinted), concrete texture.

## Anomalies and condition

Note anything that appears anomalous, damaged, incomplete, or out of place:
cracks, stains, missing parts, temporary repairs, mismatched elements,
safety hazards.

## Camera/lens artefacts

Note: lens flare, motion blur, depth-of-field blur (bokeh), barrel or
pincushion distortion, chromatic aberration, noise/grain, JPEG artefacts,
stitching seams, reflections of the photographer, timestamp overlays.

## Architecture and period

If a built environment is visible, describe the architectural style and any
period indicators: ceiling type, floor material, wall treatment, window
style, door hardware, moulding profiles, electrical fittings.

---

# Filling the fields

Apply all of the above observations to populate every field below. The depth
of content in each field should reflect the forensic standard above.

- `subject`: a concise noun phrase for the primary subject. Lower case unless
  a proper noun.
- `subject_type`: exactly one of object, person, scene, document, other.
- `objects`: list **every** discrete identifiable item, most prominent first.
  Aim for fifteen to thirty items — far more than a surface scan. Include
  materials and condition where visible.
- `scene`: the setting described in detail — location type, lighting
  conditions, time-of-day cues, weather if outdoor, room type if indoor.
  Include spatial layout: what occupies foreground/midground/background,
  camera angle, and depth cues.
- `distinctive_features`: the marks that would let someone distinguish this
  exact scene or subject from a near-identical one. Serial numbers, specific
  wear patterns, unique signage text, unusual juxtapositions, identifiable
  faces or logos. Be exhaustive — list every distinguishing element. Empty
  list only when there are genuinely none.
- `dominant_colors`: name colours with the specificity described above. List
  at least five if that many are present.
- `image_quality`: `focus`, `lighting`, `framing` each 0.0–1.0 as a
  judgement of the photograph itself. `obstructed` true when something
  blocks the main subject.
- `usable_as_evidence`: whether the photograph is clear and detailed enough
  to support a decision about the subject.
- `unclear`: true only when the subject itself cannot be confidently
  identified.
- `recommend_reshoot`: true when a retake would materially improve the
  record; `reshoot_reason` gives the single most important reason.
  Otherwise false / null.
- `tags`: lower-case words and short phrases for retrieval. Include all OCR
  text fragments, brand names, material types, activity descriptions, and
  architectural terms. Aim for fifteen to thirty tags.
- `description`: THIS IS THE MOST IMPORTANT FIELD. Write a comprehensive
  multi-paragraph forensic scene report. This must contain ALL observations
  from every section above: OCR transcriptions, people descriptions, complete
  object inventory with materials, spatial layout, brands and signage,
  lighting analysis, colour and texture specifics, anomalies and condition,
  camera artefacts, and architecture/period notes. Structure it with clear
  paragraph breaks between topics. Aim for 300-800 words — as long as needed
  to capture every observation without loss. This field is the permanent
  record; nothing observed should be omitted from it.

Be consistent between images: the same kind of subject should get the same
kind of label, because these records are grouped and compared.

Writes a new file. The version you are looking at is never modified.

Prompt profiles

forensic-v2 /app/imagelab/profiles/forensic/v2.md

Re-run under this version