Agent workflow

Draft, critique, repair, validate, render.

Agents should use Paralleliz as a deterministic slide engine: pass compact specs, inspect structured feedback, repair only what changed, then render editable PowerPoint. External LLMs can be plugged in by the caller, but the bundled loop is deterministic baseline tooling rather than hidden slide-writing magic.

Loop

5 tool-mediated steps.

MCP or Python SDK

1. Draft

Create an initial deck spec from structured analysis, requested visuals, and narrative intent.

Artifact: Draft JSON spec

Review: Slide sequence, chart choices, required fields, source notes, and template assumptions.

Validation and review tools

2. Critique

Inspect the draft against renderer constraints, missing fields, accessibility advisories, and copy risks.

Artifact: Validation report and issue list

Review: Blocking errors, warnings, accessibility notes, and over-broad editability claims.

Agent loop or local script

3. Repair

Revise the spec before rendering so the final deck is closer to the intended story and template.

Artifact: Repaired JSON spec

Review: Changed fields, normalized values, chart labels, and template placeholder fit.

CLI or service job

4. Validate

Export normalized specs, diffs, and validation reports before generating or distributing a deck.

Artifact: Normalized spec, diff, report

Review: Traceability from input to rendered output and any warnings that should block a pilot.

CLI, SDK, or HTTP service

5. Render

Generate the PPTX and hand it to a human reviewer with the supporting artifacts attached.

Artifact: Editable PPTX deck

Review: Native chart editability, shape editability, layout fit, screenshot capture readiness, and final claims.

Representative run

Founder update generated from KPI tables and narrative notes

A representative agent loop showing how Paralleliz keeps slide generation compact, inspectable, and repeatable.

Prototype

Source inputs

  • monthly_kpis.csv
  • pipeline_by_segment.xlsx
  • founder_notes.md
  • brand_theme.paralleliz.json

Validation findings

  • warning: Metric cards need source labels for reviewer traceability.
  • error: Series label is missing a stable color token.
  • advisory: Theme is supported locally; customer template mapping remains pilot-specific.

Repair diff

  • Added source note to executive summary.
  • Bound enterprise series to a stable theme token.
  • Requested review artifacts before final PPTX handoff.

Lineage

  • Inputs are kept separate from generated slide specs.
  • The agent repairs structured fields instead of rewriting the whole deck prompt.
  • Rendered artifacts keep enough context for human review and future batch indexing.

Artifacts

Compact specs, small repairs, traceable outputs.

Prompt discipline

Cache schema context, send only compact specs and validation findings, request JSON Patch for repair, then render after the diff is reviewable.

Next step: Start with the developer quickstart and a representative deck spec.

Draft spec

{
  "deck": "August Founder Update",
  "template": "operator-update",
  "theme": "paralleliz-executive",
  "slides": [
    {
      "id": "exec-summary",
      "layout": "title-metrics-narrative",
      "headline": "Revenue expanded while enterprise pipeline slipped",
      "visuals": [
        "metric_strip",
        "segment_bar"
      ]
    },
    {
      "id": "pipeline-mix",
      "layout": "chart-with-callouts",
      "headline": "SMB growth offset enterprise delays",
      "visuals": [
        "stacked_bar",
        "callout_list"
      ]
    }
  ]
}

Repair patch

[
  {
    "op": "add",
    "path": "/slides/0/source_note",
    "value": "Source: monthly_kpis.csv, closed through 2026-08-25."
  },
  {
    "op": "replace",
    "path": "/slides/1/visuals/0/series/enterprise/color",
    "value": "brand.deep-teal"
  },
  {
    "op": "add",
    "path": "/render/options/review_artifacts",
    "value": [
      "normalized_spec",
      "validation_report",
      "diff_summary"
    ]
  }
]

Render outputs

Editable deck: example_out/story_auto.pptx Implemented

Validation report: example_out/story_auto_report.json Prototype

Normalized spec: example_out/story_auto.normalized.json Planned