Conciseness

The body is two lines — an HTML comment ("Generated by scripts/agent-harnesses.py; do not edit directly.") and one instruction ("Load and follow the shared skill at `~/.agents/skills/ocr/SKILL.md`.") — with zero padding or explanation of concepts Claude already knows. Every token (aside from the trivial one-line editor comment) serves the delegation, matching the lean-and-efficient anchor; there is no over-explanation to trim, so it does not sit at 4.

Actionability

"Load and follow the shared skill at `~/.agents/skills/ocr/SKILL.md`" is a single concrete directive with a specific path, but it contains no commands, code, or steps for the OCR task itself — the actual guidance (pdftoppm/tesseract/ocrmypdf usage) is entirely absent from this file. This matches minimal-concrete-guidance anchor 2; it is not 1 because the pointer is a real, specific instruction rather than vague abstraction, and not 3 because no partial task guidance exists here.

Workflow Clarity

The described task is a multi-step pipeline (render PDF pages via pdftoppm, run tesseract, fall back to ocrmypdf, verify output), but the body provides no sequence, steps, or validation for any of it — the workflow lives entirely behind an external pointer. This matches anchor 1 ("Steps missing... no sequence; no validation"); the simple-skill exception does not apply because OCR is multi-step, and even the delegation target is a dead end in this bundle.

Progressive Disclosure

The body's only reference is clearly signaled and one level deep, but it points outside the skill bundle to `~/.agents/skills/ocr/SKILL.md`, which does not exist in this environment, and the bundle ships no references/, scripts/, or assets/ files at all — so the skill contains no navigable content, only a broken external pointer. That lands at minimal-structure anchor 2 rather than 3, since the failure is not organization of inline content but a complete absence of reachable material.

Specificity

"OCR scanned PDFs and image files into searchable text using pdftoppm + tesseract (or ocrmypdf if available)" names concrete inputs (scanned PDFs, image files), a concrete output (searchable text), and specific tools including a fallback — several specific elements with only minor gaps. It falls short of 5 because it describes a single conversion action rather than multiple distinct actions, but sits above 3 since the toolchain and I/O are explicitly specified.

Completeness

It explicitly answers both questions: what ("OCR scanned PDFs and image files into searchable text using pdftoppm + tesseract") and when ("Load when about to OCR a scanned PDF or image, or when extracted PDF text is empty/garbled") with concrete trigger phrases including the subtle empty/garbled-text case. This matches the anchor-5 example structure; anchor 4 would require the 'when' to be less explicit, which it is not.

Trigger Term Quality

Natural phrases users would say are present: "OCR", "scanned PDF", "image", "extracted PDF text is empty/garbled", plus the tool names tesseract/ocrmypdf users might mention directly. Missing a few natural terms such as file extensions (.pdf, .png, .jpg) and synonyms like screenshots or photos, so it does not reach comprehensive 5 coverage.

Distinctiveness Conflict Risk

OCR is a clear niche with distinct triggers (scanned PDFs, garbled extracted text) unlikely to fire for general PDF-editing or text-extraction skills — the "extracted PDF text is empty/garbled" condition actually delimits the boundary against plain extraction skills rather than overlapping them. Minor adjacency to PDF-extraction skills exists, but not enough to drop to 4.