Case studies · Retrospective evidence
These are documented AI-assisted project cases. For product, marketing and operations assignments, see the business work samples.
What was asked. What was checked. What remains unknown.
The work becomes credible when the reader can follow the decision.
Robert set project direction and acceptance constraints. AI agents assisted analysis, implementation and verification. The summaries identify project decisions, AI assistance and recorded checks.
Sovereign Mercy · PennedCanon · The Sage Dad · Stangl Market
Sovereign Mercy: revising without losing the story
October 2026 · Author-directed, AI-assisted editorial and production workflow
The problem and judgment
A reader reported too much dialogue while reading Chapter 5. Robert asked whether the problem extended further, then authorized a targeted revision: preserve canon, Mara’s role and the length of the intimacy scene, with review before publication.
Dialogue volume was a diagnostic clue. A lower percentage would not itself prove better writing. The review distinguished repetitive explanation from consequential disagreement, investigation and relationship development. Robert subsequently approved the revisions and specified which publication systems should change.
The method
- Retain the source edition and compare proposed changes chapter by chapter.
- Check protected chapters and passages, scene breaks and story constraints.
- Keep the review copy separate from the approved edition.
- After approval, check exact prose in live or queued chapters and EPUB packages.
- Preserve release dates, commercial settings and earlier editorial decisions.
What the retained evidence establishes
| Check | Recorded result | Limit |
|---|---|---|
| Protected prose | The October 21 dialogue-pass checks record Chapter 14 and two protected Chapter 13 passages as unchanged. | Preservation does not establish literary quality. |
| Publication destinations | 22 chapter checks record exact prose at the intended live or queued destinations. | Dated evidence, not a new inspection of today’s production state. |
| EPUB contents | 10 chapters in Part One and 9 in Part Two recorded as text-verified. Both EPUBCheck logs report zero errors and warnings. | Format validity differs from reader satisfaction and device rendering. |
| Release control | The final check records unchanged queue dates and status, prices and classification. | No claim of current queue status or an error-free system. |
A useful imperfection
The intermediate reconciliation record retained a citation on its original version when wording changed. A later manual-reconciliation note and final snapshot document further work. Other unresolved records and an analytics-refresh failure remain in the historical snapshot. That evidence does not support a claim that every system passed.
Robert’s contribution: editorial problem, scope, protected content, acceptance decision and publication constraints. AI assistance: analysis, proposed edits, comparison tools and recorded checks.
Inspect the evidence digest → · Explore the published work →
PennedCanon: making AI assistance understandable
October 2026 · Product direction and usability QA
The problem and intervention
AI assistance was buried in the scene editor. Robert asked for consistent treatment of AI actions and explicitly limited the request to presenting existing capabilities. He did not authorize building imagined sentence-level or character-brainstorming features in that pass.
The recorded update introduced an “AI assistance · This scene” entry point. It distinguished built-in drafting and revision, “Package for another AI · Culverin,” and previous AI drafts. The label stated the action’s scope.
What was checked
The implementation chat reports desktop and 390-pixel phone-width checks, opening the existing Culverin workflow, menu dismissal and keyboard access. It reports no AI generation or manuscript changes during testing. Physical iPhone testing was expressly unverified.
Robert’s contribution: identifying discoverability and consistency problems, defining scope and challenging the presentation. AI assistance: implementing the UI and carrying out reported checks.
Outcome: a more explicit entry point for existing functions, according to the retained implementation record. Task-completion time, adoption and error reduction were not measured.
The Sage Dad: investigating a warning before changing access
October 22, 2026 · Agent-assisted operational diagnosis
A Search Console notice reported access-forbidden pages. The investigation needed to establish whether public resources were affected or book downloads were intentionally restricted.
The retained inspection identified four download URLs. Public activity pages described book-code companions not yet available to unlock. The sitemap contained no download URLs, while the public markup contained hidden download anchors. The recommendation was to preserve access restrictions and avoid exposing restricted links before access was granted.
Evidence limit: an agent-produced diagnosis within Robert’s operating workflow. It did not establish Google’s historical discovery path, test every endpoint or demonstrate a shipped remediation. Conflicting report dates were disclosed.
Evaluation relevance: define expected behavior before labeling a defect; distinguish a symptom from its possible cause; avoid a “fix” that breaks a legitimate restriction.
Stangl Market: product boundaries that matter
The existing public case study documents vendor applications, organizer review, profiles and weekly lineups. It distinguishes accepting an application from publishing a profile or confirming a date, and creating an email snapshot from sending a newsletter.
These examples connect product judgment to testable state changes. The case study documents delivered functionality; attendance, sales, traffic growth and time savings remain unmeasured.
Further evaluation
The proposed protocol outlines a controlled comparison of single-agent and multi-agent workflows, with separate measures for quality, correction effort, time and cost.