Lean and assumes Claude's competence — no explanations of what an API or idempotency is; every line earns its place (key-loading gotcha, field limits, failure-mode recovery). Not the verbose score-1 anchor and tighter than the score-2 'good library' padding.

Provides fully executable, copy-paste bash/jq/curl snippets (uuidgen, jq --rawfile, curl with headers) plus concrete jq parse commands — matching the 'fully executable code/commands' anchor, not pseudocode.

Clear three-step sequence (build prompt → run → read report) with an explicit validation checkpoint ('jq -r .status' succeeded|failed) and error-recovery feedback loops (402 top-up + replay-safe retry, 429 Retry-After, 502 report-and-stop), exceeding the score-2 'steps listed but checkpoints missing' anchor.

No bundle files exist, and this single-file skill is well-organized into focused sections with one-level references out to the deepapi and research-prompt skills; per the simple-skill scoring note, well-organized single-file content scores 3 here.

Lists multiple concrete actions — 'Builds a rigorous one-paragraph research prompt', 'fires it', 'saves a cited markdown report', 'POST /v1/research/deep' — matching the 'multiple specific concrete actions' anchor rather than the partial score-2 list.

Explicitly answers both what (build prompt, fire, save cited report) and when ('Use when the user asks for ...'), with an explicit trigger clause, so it clears the completeness cap of 2.

Covers natural user phrasings — "deep research", "deepapi research", "perplexity deep research" (legacy), and 'deep source-backed research run' — giving good coverage of terms a user would actually say, not just jargon.

Explicit differentiator ('this is the full research workflow ... not raw endpoint access') plus distinct deep-research triggers carve a clear niche unlikely to fire for the sibling deepapi skill.