From drawing observations to checkable scoring and usable reports

Support institutional use, beyond generating a report

An institution needs to collect material, run analysis, review results and manage reports. Participants provide input; professionals interpret individual results; administrators coordinate a group of tasks and subsequent work.

I defined the product around the complete path from input to use:

1. 01Collect or upload drawing
2. 02Analyse visual features
3. 03Apply scoring rules
4. 04Organise results and interpretation
5. 05Generate report
6. 06Manage institutional results

The individual path needs to explain one result. The batch path needs to organise multiple records, processing states and reports. Both reuse the analysis rules but serve different operating tasks. This meant designing institutional workflows and result management alongside the prompts.

Turn specialist interpretation into explicit analysis requirements

“Observe the composition” or “analyse signs of pressure” is not a checkable AI task. I needed to specify the object, visible conditions, scoring relationship and where the result would appear.

My work connected three levels: translating specialist material into visual features and rules, specifying model analysis, and designing how people inspect and use the results. Interpretive meaning stays separate from judgments about what is visibly present in the drawing.

2.1 From feature recognition to report interpretation

The rule documents define features, scoring, combined judgments and report requirements. The analysis prompts specify their order. The following describes logical responsibilities and intended outputs; technical deployment is a separate question.

These are logical steps in the rules and prompts, rather than proof of separately deployed models or calculation services. Keeping that distinction clear also helps locate differences in recognition, calculation or explanation.

2.2 House–Tree–Person: classify the feature before defining the condition

I organised the rules into overall composition, house, tree and person, then into observations such as position, size, lines, structure, environment and posture.

Overall composition covers size, position, clipping, perspective, lines and form. House covers proportions, structure, viewpoint, style and surroundings. Tree includes clipping, tilt, crown–trunk proportions, form and season. Person covers appearance, orientation, posture and proportions.

Each detailed rule has five parts:

1. 01Feature code
2. 02Description
3. 03Visible decision condition
4. 04Scoring contribution
5. 05Interpretation reference

For example, subject size is defined relative to the area of the paper. Person size uses height relative to paper height; tree proportions compare crown and trunk. Each object needs its own reference, rather than one vague definition of “large.”

The visible definition establishes what is present. The interpretation reference describes how that observation is used within the assessment scheme. The document separates them so they can be discussed and revised independently.

2.5 Person-in-the-Rain: separate pressure, emotional expression, resources and warnings

I did not reuse the House–Tree–Person object structure unchanged. Person-in-the-Rain uses three groups and a separate warning check:

• A — Perceived pressure: rain intensity, drops, clouds, lightning, standing water and getting wet.
• B — Emotional expression: expression, posture, size, lines, shading and body-related features.
• C — Coping resources: umbrella, raincoat, shelter and other resource-related features.
• Special warnings: checked separately in the combined judgment and report, rather than added directly to the dimension sum.

This retains pressure-related and resource-related observations together, rather than reducing a drawing to one good/bad label. The rules specify how A and B form the combined pressure indicator, how C enters the calculation, and how warning conditions affect the attention level.

2.7 Write prompts as execution requirements

Alongside the rules, I specified analysis order and output requirements. Person-in-the-Rain starts with rain, the person and core protective elements, then moves to auxiliary elements. With multiple people, the main person is identified first and other people are considered through the relevant relationships.

Numerical processing records item results before dimension and combined calculations, retaining raw results. Warning checks remain a separate step; generating a combined score does not finish the analysis.

Ambiguous and exceptional inputs have their own requirements. Uncertain local elements are marked as such. If core content cannot be recognised, the output explains that it cannot be scored. Special conditions stay in the report instead of being invented by the model.

These requirements specify sequence, decision boundaries and stopping conditions: when to continue and when a complete result is inappropriate.

Separate recognition variation, calculation differences and inconsistent explanation

Outputs include feature judgments, dimension results and report explanations, not just a total score. When repeated analysis of the same image differs, the question is where the difference occurred: recognition, rule application or explanation.

I repeatedly analysed the same drawing to compare recognition, scoring and interpretation. I also tested image rotation, camera tilt and changes in lighting to examine behaviour under different input conditions.

Inspect recognition, scoring and reporting separately

Recognition: is the model judging the same condition?

I compared recognised features against the rule's visual conditions. Similar-looking prose is not enough: subject size has a paper-area reference, while the umbrella rule distinguishes absence, ineffective cover and effective cover. The check concerns how the model applies those conditions and whether input presentation changes that judgment.

Scoring: how does recognition affect the result?

I checked correspondence between features, scoring conditions and final results across repeated runs. Item contributions, dimension membership, cumulative limits and combinations provide different places to investigate. A different total alone does not explain the cause.

Reporting: does the explanation correspond to the result?

I checked that explanations were clear, followed the rules and corresponded to recognition and scoring. Report generation is not an unrelated writing task. Features, dimensions and flags need matching explanations; fluent prose alone does not establish that connection.

Test input conditions as a separate question

Beyond repeated analysis of an unchanged image, I tested rotation, camera tilt and lighting. Repetition checks consistency for the same input; altered presentation checks sensitivity to input conditions. Keeping them separate helps locate what needs further investigation.

Refine three connected parts

Recognition and scoring prompts

I revised visual conditions and output requirements to make tasks more specific and reduce reliance on broad descriptions.

Rules and weights

I adjusted feature conditions and weight configuration, refining how recognition entered scoring. Rules define the judgment and contribution; prompts ask the model to apply them.

Analysis workflow

I adjusted analysis steps and processing logic to connect recognition, scoring and output, then repeated checks. Prompts, rules and workflow are interdependent, so a changed output cannot automatically be attributed to one technical change.

My contribution and the evidence boundary

I participated in testing, comparison and refinement while writing both assessments' scoring rules and analysis prompts. This connected specialist rules, model tasks and result interpretation in one product workflow.

The diagnostic approach is to check visible features first, rule use second, and correspondence in the report third. The available public material does not include a complete sample-level before/after record or a measured improvement rate; I do not claim a proven performance gain.

Individual: organise information around one result

The individual path connects upload, analysis, result and report. Professionals need both the overview and dimension explanations. The drawing, identity details, assessment date and result form one record.

The Person-in-the-Rain report template retains the drawing, basic details, dimensions and combined score, then expands pressure, emotions, resources, warnings and conclusions. Readers can navigate by theme and return to related items.

Batch: organise records around an institutional task

Batch processing connects multiple inputs, processing states, result review and report management. I designed individual/batch analysis, analysis records, result review, report management and knowledge management for institutional workflows, reusing the same analysis capability at different operating scales.

Customisation: balance shared analysis with institutional differences

Institutions differ in assessment organisation, result review and report delivery. I worked on requirements, prototypes and workflows, supporting 20+ client-specific versions. These were customised deliveries, not a claim of 20 standard-product iterations.

My responsibilities included product planning, rules and prompts, prototypes and platform workflows, and coordination with engineering. Rules and report requirements describe the analysis; specific feature trade-offs need to be understood in the relevant institutional context.

The product has supported 7,000+ AI assessments, with 160+ visual feature dimensions and 20+ client-specific versions, entering education, healthcare and other institutional settings.

My main learning was how to make specialist content discussable and checkable: give observations codes, define how recognition enters scoring, specify the report's evidence, and connect these capabilities to individual, batch and administrative workflows.

Clear rules do not prove stable recognition. A score does not prove that the explanation corresponds. A report does not prove that an institution can use it smoothly. Product work connects these steps individually.