AI Skill Report Card

AI Automation Overlay for Data To Product Repackaging

C+58·Sep 26, 2026·Source: Extension-selection

AI-Automation Overlay for Data-to-Product Repackaging

10 / 15

Reframe every atomic block from the original 37-block pipeline as a typed function: block(input_schema, confidence_threshold) → {output_schema, confidence, logs}. Chain these into a DAG orchestrated by a planner agent. Route low-confidence outputs to human review. This skill converts a manual repackaging workflow into an agentic automation pipeline.

classify_source(file) → {source_type, modality, structure_schema, parser_route, confidence}
  if confidence < 0.8 → escalate_to_human()
  else → route_to_parser(parser_route)
Recommendation▾
Missing YAML frontmatter entirely (no 'name' or 'description' fields) — this is a critical format violation that should be added immediately for the skill to be discoverable/loadable.
13 / 15
Progress:
- [ ] Step 1: Define orchestrator + shared memory (vector DB, metadata store, asset store, job queue)
- [ ] Step 2: Map each atomic block to an agent/function with I/O schema
- [ ] Step 3: Assign AI capability + tooling per block
- [ ] Step 4: Set confidence thresholds and HITL gates
- [ ] Step 5: Wire blocks into DAG by phase dependency
- [ ] Step 6: Execute, log, and route exceptions
- [ ] Step 7: Close feedback loop into re-triggering upstream blocks

Phase 0 — Orchestration Layer (build this first)

  • Shared memory: vector DB (embeddings), metadata store, asset store, version control, job queue.
  • Agent roles: Source Scout, Rights Auditor, Quality Scorer, Atomizer, Context Reframer, Metadata Tagger, Transformer, Assembler, Renderer, QA Agent, Publisher, Valuator.
  • Execution pattern: event-driven DAG; every node = input_schema → function → output_schema + confidence + logs.
  • Escalation rule: confidence < threshold, legal ambiguity, brand risk, or high print-run cost → human gate.

Phase 1 — Source Acquisition (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
1.1 Source Type Classificationclassify_source(file) → {type, modality, schema, route}VLM/LLM + file sniffer + OCRconfidence < 0.8 or unknown format
1.2 Rights & Licensingaudit_rights(source) → {status, license, allowed_derivatives, risk_flags}Legal RAG + license classifierany risk_flag = true
1.3 Source Qualityscore_source(source) → {completeness, formatting, metadata_density, fixes}Heuristics + LLM-as-judgescore below threshold
1.4 Source Aggregationaggregate(targets) → {corpus, dedupe_map, embeddings, manifest}Crawlers/connectors + MinHash + embeddingsscope/ToS approval

Phase 2 — Atomization & Extraction (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
2.1 Idea Identificationextract_atoms(text) → [{type, text, location, embedding}]LLM structured extraction, semantic chunkerprecision/recall spot-check
2.2 Unit Qualificationqualify_atom(atom) → {standalone_score 1-5, reason}LLM-as-judge / reward modelscore ≤ 3
2.3 Context Repatriationrepatriate(atom) → {atom+context, entailment_score}LLM generation + NLI verifierlow entailment or high-stakes domain
2.4 Element Metadata Taggingtag_element(element) → {type_tags, confidence}Zero-shot classifier / ontology taggermis-tag correction
2.5 Format Inventoryinventory(assets) → {format_matrix, missing, conversion_tasks}File-type detector + code interpreterconversion priority approval

Phase 3 — Transformation & Enrichment (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
3.1 Pretextual Wrappergenerate_wrapper(source, atoms) → {intro, commentary, theme}LLM + RAGeditorial voice review
3.2 Cross-Platform Adaptationadapt_channels(atom) → {email, web, social, sms, print}LLM + constraint enginebrand/compliance review
3.3 Format Translationtranslate_format(text) → {faq, list, table, qa}LLM structured outputfidelity check
3.4 Style & Brandingapply_brand(content, kit) → {styled_assets}LLM + image gen + CSS enginebrand approval
3.5 AI Script Enrichmentenrich(data) → {summaries, indexes, entities, links}Code interpreter + RAG + NERaccuracy validation

Phase 4 — Assembly & Packaging (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
4.1 Content Pyramidplan_pyramid(pillar) → {15–25 assets, formats, deps}LLM planner + dependency graphplan approval
4.2 Template Mappingmap_template(atoms, templates) → {filled_templates}LLM + template enginelayout exceptions
4.3 Modular Reassemblyreassemble(atoms, rules) → {configurations}Constraint solver + LLMnovel config approval
4.4 Sequencingsequence(atoms) → {toc, order, rationale}LLM + graph rankingeditorial override
4.5 Cross-Referencinglink_atoms(atoms) → {internal_links, anchors}Embeddings + LLMlink relevance check

Phase 5 — Digital Download Generation (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
5.1 PDF Compilationcompile_pdf(content) → {pdf, cover, toc}DOMPDF/Prince/Pandoc + AI cover genfinal proof
5.2 EPUB Constructionbuild_epub(source) → {epub, spot_check}Pandoc + epubcheck + LLM translationstratified spot-check
5.3 Gated Packagingpackage_gated(product) → {optin_page, delivery_flow}LLM copy gen + form APImarketing approval
5.4 Marketplace Listinglist_product(product) → {copy, pricing, tags}LLM + market data APIpublish approval

Phase 6 — Printable Product Generation (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
6.1 Data Prepclean_data(csv) → {clean_data, schema, anomalies}Code interpreter + LLMschema change confirm
6.2 Variable Template Designdesign_template(schema) → {template+fields}InDesign scripting / Canva APIdesign approval
6.3 Merge & Preflightmerge_preflight(template, data) → {pdfs, report}Data Merge + preflight toolsfix flagged issues
6.4 Imposition & Batchimpose(pdfs) → {press_ready}Imposition software + scriptsphysical proof
6.5 Print-Ready Exportexport_print(imposed) → {pdfx, upload}PDF/X validator + print APIfinal print approval

Phase 7 — Post-Production & Distribution (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
7.1 QAqa(outputs) → {defects, severity, fixes}VLM (layout) + LLM (content)critical defect
7.2 Metadata Embeddingembed_metadata(product) → {title, rights, a11y_tags}LLM + XMP/EPUB toolsrights verification
7.3 Version Controltrack_versions(changes) → {diffs, rollback}Git/DVC + media APImajor version approval
7.4 Distribution Setupdistribute(product, channels) → {listings, uploads}Marketplace APIs + RPAcompliance confirm
7.5 Feedback Loopoptimize(perf_data) → {revision_tasks, ab_tests}LLM + analyticsprioritization review

Phase 8 — Estimation & Valuation (AI-Reframed)

BlockAI TaskModel/ToolHITL Trigger
8.1 Cost Approachestimate_cost(source) → {replacement_cost}Regression + LLMassumption validation
8.2 Income Approachforecast_income(product) → {dcf, projections}ML forecastingprojection review
8.3 Market Approachfind_comps(product) → {comps, valuation_range}Embeddings + market DBcomp confirmation
8.4 Max Size Estimationestimate_size(content) → {uncompressed, compressed, page_count}Code interpreter + compression toolsprint-run approval
Recommendation▾
The skill assumes familiarity with an external '37-block pipeline' that isn't defined here, making it incomplete as a standalone artifact; either inline the block definitions or clarify this is a companion/overlay skill with an explicit dependency.
Scout→Rights→Quality→Aggregator
  → Atomizer→Qualifier→ContextReframer→Tagger→FormatInventory
  → WrapperGen→ChannelAdapter→FormatTranslator→BrandStyler→Enricher
  → PyramidPlanner→TemplateMapper→Assembler→Sequencer→Linker
  → [PDF/EPUB Compiler→Gated Packager→Marketplace Lister]
    OR [DataCleaner→TemplateDesigner→Merge/Preflight→Imposition→PrintExporter]
  → QA→MetadataEmbedder→VersionController→DistributionAgent→FeedbackOptimizer
  → CostValuator→IncomeForecaster→MarketValuator→SizeEstimator
  • Every block gets a confidence score — never let low-confidence output flow downstream silently.
  • Keep human gates at: rights ambiguity, brand approval, high-cost print runs, final publish, financial assumptions.
  • Use RAG/embeddings for anything requiring grounding (rights, comps, context repatriation) — pure generation hallucinates here.
  • Version every atom independently; don't rebuild the whole product to fix one block.
  • Skipping the rights-audit gate to save time — creates legal exposure that automation can't undo.
  • Letting the LLM self-score qualification (2.2) without an independent judge — inflates standalone-strength ratings.
  • Automating print imposition (6.4) without a physical proof step — batch errors scale to the full print run.
  • Treating valuation blocks (Phase 8) as deterministic — always output a range with confidence, not a single number.
0
Grade C+AI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
10/15
Workflow
13/15
Examples
10/20
Completeness
16/20
Format
6/15
Conciseness
10/15