| name | component-split-ocr-review |
| description | Export deterministic component-level OCR evidence for one transparent or semi-transparent image by splitting alpha-connected components, writing a component table, and running OCR on each component. Use when bounded component evidence must be exported before any downstream refinement or caption rerun. |
Component Split OCR Review
Overview
Use this skill to turn one local image into a bounded review surface:
- alpha-connected component crops
- component summary table
- per-component OCR evidence
- markdown and JSON artifacts for human review
This skill wraps the shared component split/OCR builder. It does not own semantic object selection, imagegen repair, or caption regeneration.
Use This Skill When
- one local image already exists
- the image is transparent or semi-transparent, or at least worth checking for alpha-separated parts
- a human needs to inspect component-level slices before deciding what to keep
- OCR on each separated component is useful evidence
Do Not Use This Skill When
- the task is full-image caption generation
- semantic object isolation or model-assisted correction is the current task
- batch subset triage across many images is the current task
- final image approval or rename/metadata commit is the current task
- the underlying tool lifecycle is broken — route to
vendored-mcp-onboarding
Required Inputs
- one absolute or repo-relative
--image-path
- optional
--output-root
- optional
--alpha-threshold
- optional
--min-pixels
- optional
--padding
- optional
--min-components-for-success
Script
scripts/build_component_split_ocr_report.py
References
knowledge_bases/component-split-ocr-review-knowledge-base-at2026-03-28-01-20.md
references/runtime.md
references/troubleshooting.md
Workflow
- Confirm the target is one bounded local image.
- Run the skill-local wrapper, not a handwritten ad hoc command.
- Let the shared builder perform alpha split first.
- Use the generated component table and per-component OCR as the review surface.
- Treat the output as review evidence only until a human decides what components are meaningful.
- Promote only reviewed components into later object-isolation or caption flows.
Preferred Output Surface
- default output root:
control/project_domain/archive/component_split_ocr/
- one bounded subdirectory per source dataset and image stem
- keep review artifacts in
project_domain/runs, not user_decisions, until a decision is explicitly promoted
Outputs
COMPONENT_SPLIT_OCR_REPORT.md
COMPONENT_SPLIT_OCR_REPORT.json
alpha_components/*.png
component_ocr/*_OCR_RESULT.json
Known Good Fit
- transparent PNGs with disconnected visual parts
- slide-export images where alpha-separated icons or panels need inspection
- weak semantic isolation cases where deterministic component evidence is needed first
Not Owned Here
- semantic object selection
- imagegen correction
- full-image caption rerun
- batch triage across directories
- final approval or commit
- component evidence를 multimodal loop state로 재주입하거나 refinement할 때 →
multimodal-evidence-refinement-loop
- review surface 정규화 및 manifest split이 필요할 때 →
image-text-cot-review
- tool lifecycle integrity (launcher, registration, inventory, setup state)