| name | content-analysis-pro |
| description | Production-grade multimodal content analysis: explicit analysis pipeline (modality → decode → segment → extract → verify → report), evidence and provenance rules (page, time, region anchors), grounded vs inferred claims, failure modes (OCR error, sampling gaps, chart number invention, deepfakes, token limits, locked files), decision trade-offs (summary vs extract vs compare, full read vs stratified sample, human-in-the-loop for high-stakes), quality and anti-hallucination guardrails, structured reports with limitations and confidence — for documents, images, video, and audio. Not a replacement for legal, medical, or forensic experts.
Use when the user supplies or points to content to summarize, extract, compare, or audit with traceable evidence. Combine with business-analysis-pro for BRD-style outputs, security-pro for PII/secrets, data-analysis-pro for tabular math on extracted data, web-research-pro for external fact-check, image-processing-pro for raster prep, testing-pro for extraction-regression tests.
Triggers: "analyze this image", "analyze video", "analyze PDF", "summarize attachment", "OCR", "transcript", "screenshot", "extract from doc", "compare these files", "timeline", "evidence map", "provenance", "grounding", "hallucination", "illegible", "frame at timestamp", "password protected PDF", "deepfake", "sampling strategy", "entity extraction", "redact", "multimodal", "chart in image", "quote with page number".
|
| metadata | {"short-description":"Content analysis — pipeline, evidence, failure modes, grounding, reports","content-language":"en","domain":"content-analysis","level":"professional"} |
Content analysis (professional)
Skill text is English; answer in the user’s preferred language when rules or the conversation specify it.
Use this skill when source material (files, pasted text, described media) must be read carefully and returned as structured intelligence with traceable evidence — not casual chat. This skill encodes a pipeline system model (analysis-pipeline-system-model.md), failure modes, decision trade-offs for coverage vs cost, and grounding guardrails. Optional tooling (FFmpeg, OCR APIs) implements stages — analysis-methods-and-frames.md — the skill governs what is asserted and how limitations are disclosed. Confirm user goal, output format, and sensitivity (PII, confidential).
Boundary
content-analysis-pro owns reading, structuring, synthesizing, and evidence-linked reporting on provided content across modalities. business-analysis-pro owns business requirements packaging from findings. data-analysis-pro owns quantitative analysis of tabular data (statistics, pivots), typically after clean extraction. security-pro owns handling policies for secrets seen in screenshots. image-processing-pro owns pixel transforms — not semantic interpretation alone.
Related skills (this repo)
| Skill | When to combine with content-analysis-pro |
|---|
business-analysis-pro | Turn extracted facts into requirements, BRD, decisions |
security-pro | Redaction, credentials/sensitive handling |
data-analysis-pro | CSV/Parquet numeric work on structured extracts |
web-research-pro | External verification of claims not in corpus |
image-processing-pro | Raster prep before vision read |
testing-pro | Golden tests for automated extraction pipelines |
repo-tooling-pro | Shared scripts for repeatable FFmpeg/OCR |
seo-pro | Rare — web/marketing asset angle |
When to use
- Documents, images, video (or paste) requiring detail, extraction, comparison, or timeline.
- Formal-style reports with sections, citations (page/time), confidence.
- Explicit limitations and uncertainty when media is ambiguous.
When not to use
- Pure tabular analytics without narrative source reading —
data-analysis-pro may lead if data already clean CSV.
- Choosing business priorities without document —
business-analysis-pro.
Required inputs
- Goal (summarize / extract / compare / compliance scan), modalities, constraints (length, language, redaction).
Expected output
Follow Suggested response format strictly — pipeline through residual risks — with evidence anchors unless user explicitly waived (then state waiver).
Workflow
Apply Karpathy principles throughout: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution.
- Confirm goal, modality, sensitivity, evidence granularity → verify: [scope defined].
- State assumptions about content type, analysis depth, evidence needs (Think Before Coding).
- Apply minimum analysis first; add depth only when justified (Simplicity First).
- Make surgical changes — only analyze content directly related to the request (Surgical Changes).
- Define success criteria (evidence quality, coverage, accuracy); loop until verified (Goal-Driven Execution).
- Respond using Suggested response format; delegate BRD packaging to
business-analysis-pro when appropriate.
Operating principles
- Think Before Coding — State assumptions: content type, analysis depth, evidence needs. Ask when uncertain.
- Simplicity First — Start with basic analysis; add depth only when justified.
- Surgical Changes — Only analyze content directly related to the request.
- Goal-Driven Execution — Define evidence quality, coverage, accuracy targets upfront.
- Ground in the source — Cite page, timestamp, or visible region; separate quote from interpretation —
evidence-and-provenance-rules.md.
- No fabricated facts — Unclear → say so; inference labeled —
failure-modes-detection-mitigation.md.
- Structured output — Headings, bullets, tables; executive + detail when useful —
reporting-and-limitations.md.
- Modality-aware — Scanned PDF, charts, video sampling —
analysis-methods-and-frames.md.
- Safety and privacy — Minimize secret reproduction —
security-pro.
- Human-in-the-loop for high-stakes domains — assist only —
reporting-and-limitations.md.
Analysis pipeline — system model (summary)
Stages from input to evidence-backed report — analysis-pipeline-system-model.md.
Details: references/analysis-pipeline-system-model.md
Failure modes — detection and mitigation (summary)
OCR, sampling gaps, hallucinated labels, DRM lock — failure-modes-detection-mitigation.md.
Details: references/failure-modes-detection-mitigation.md
Decision framework and trade-offs (summary)
Coverage vs depth; human review domains — decision-framework-and-tradeoffs.md.
Details: references/decision-framework-and-tradeoffs.md
Evidence and provenance rules (summary)
Anchors per modality; quote discipline — evidence-and-provenance-rules.md.
Details: references/evidence-and-provenance-rules.md
Quality validation and grounding guardrails (summary)
Anti-hallucination checklist — quality-validation-and-guardrails.md.
Details: references/quality-validation-and-guardrails.md
Analysis methods and frames (summary)
Per-modality angles; video paths — analysis-methods-and-frames.md.
Details: references/analysis-methods-and-frames.md
Reporting and limitations (summary)
Report sections, fidelity, confidence — reporting-and-limitations.md.
Details: references/reporting-and-limitations.md
Tips and tricks (summary)
Goal-first, sampling, scans — tips-and-tricks.md.
Details: references/tips-and-tricks.md
Edge cases (summary)
DRM, tables, RTL, litigation sampling — edge-cases.md.
Details: references/edge-cases.md
Scope and file-format dispatch (summary)
Routing formats — file-formats-dispatch-and-scope.md.
Details: references/file-formats-dispatch-and-scope.md
Decision trees (summary)
Input type, goal, scale, risk — decision-tree.md.
Details: references/decision-tree.md
Anti-patterns (summary)
No sampling frame, OCR blindness — anti-patterns.md.
Details: references/anti-patterns.md
Integration map (summary)
business-analysis-pro, security-pro, data-analysis-pro, web-research-pro — integration-map.md.
Details: references/integration-map.md
Versions and tooling (summary)
Tool/docs drift — versions.md.
Details: references/versions.md
Suggested response format (STRICT — implement / review)
- Context — What was supplied (modalities, count, language hints), user goal, sensitivity level.
- Problem — Deliverable shape (summary vs extraction schema vs comparison) and success criteria.
- System design / architecture — Pipeline sketch: modalities → segmentation/extraction → synthesis; grounding rule for this run —
analysis-pipeline-system-model.md.
- Decision reasoning — Coverage strategy (full vs sample), frames (entities/timeline), handoff to
business-analysis-pro / data-analysis-pro if needed — decision-framework-and-tradeoffs.md / decision-tree.md.
- Implementation sketch — Report outline, extraction table schema, or section list — structured body (label Code if using a fixed template block).
- Trade-offs — Depth vs token/time; automation vs manual verify; multilingual risk.
- Failure modes — What could mislead (OCR, missed scene, illegible chart) —
failure-modes-detection-mitigation.md themes.
- Residual risks — Uncertainty pockets, need for human expert,
security-pro if secrets surfaced.
Resources in this skill
Quick example
Input: Dashboard screenshot — “Is revenue growing?”
Expected output: Chart read from visible axes; confidence; no invented numbers; failure modes if unreadable.
Input: Password PDF — extract vendors.
Expected output: Cannot without unlock; no guessed content; security-pro-safe path.
Input: Webinar → acceptance criteria.
Expected output: This skill: timestamped summary; business-analysis-pro: AC; demo ≠ contract risk in residual.
Checklist before calling the skill done
Grounding
Process & handoffs
Robustness
