Published case study — attributed implementation
saved.mdReports, proposals and presentations
From AI chat to a shareable web page.
- The need
- Professionals draft useful work with AI, then copy it into another tool to format and share. Sending the whole chat exposes the working conversation instead of giving a client or colleague the finished result.
- What we built
- We built a publishing platform that turns selected chat content into web pages, reports and interactive presentations. Users approve publication and updates, with version history and reusable templates.
Illustrative stock photo, not project imagery.alleksana / Pexels
Published case study — attributed implementation
This page documents only the structure or implementation supported by the recorded sources. No measured outcome or testimonial is currently approved for publication.
The published web experience
Explore selected screens from the project. The complete site opens separately for current content and full interaction.
The operational problem
An AI conversation is a productive workspace, but it is a poor final format for client or internal delivery. Copying the result into another tool loses time, structure and interactive behaviour. Sharing the chat exposes too much context and makes the reader work through the drafting process instead of seeing the finished outcome.
This page separates implemented structure from unverified outcomes or claims. Existing public identities remain bounded by each case's declared status and sources. Any new or expanded identity, testimonial, metric, or verified outcome requires claim-specific evidence and publication approval.
Current-state flow
The repeated handoffs the system is designed to replace or make visible.
- 01
A professional asks an AI tool to turn the current conversation into a defined page or document.
- 02
The web agent shapes the relevant context into Markdown, HTML or JSX and prepares a saved.md publication action.
- 03
The user reviews and approves the action before the page is published.
- 04
The finished deliverable receives a clean web link and can be updated, versioned, remixed or exported when the work changes.
Constraints and risks
These conditions shape the architecture before automation begins.
Publication needs clear consent
Conversation context can contain drafts or private details, so the user must choose what becomes part of the page and approve the publishing action.
Several page formats must render reliably
Markdown, HTML and JSX have different capabilities. The platform needs predictable rendering, safe boundaries and a usable result across desktop and mobile screens.
Updates cannot overwrite newer work
A page may be edited by a person or agent after publication, so version checks must protect the latest approved state from stale changes.
System design
A traceable path from intake to action, with uncertainty surfaced before it becomes an operational error.
Agent publishing interface
An MCP connection lets supported AI tools create, retrieve and update saved.md pages from the conversation where the work begins.
Multi-format page renderer
The publishing layer turns Markdown, HTML or JSX into a hosted web page, including light interaction and presentation controls where the format supports them.
Page lifecycle
Ownership, approval, version history, reusable templates, remixing and export keep each deliverable useful after its first publication.
Responsibility by design
The system does not treat every task as an AI task.
Deterministic responsibilities
Repeatable rules own validation, state, and system changes.
- Validate the requested page format and required publication fields.
- Create a stable page record and a public link with an opaque identifier.
- Render Markdown, HTML and JSX through the correct page runtime.
- Retain the current page version and reject updates based on stale state.
- Preserve page ownership and collaborator editing permissions.
- Generate a PDF export when the user needs a file version of the published page.
Bounded-AI responsibilities
AI assists with narrow interpretation work and exposes its basis.
- Select the relevant context from the conversation and leave unrelated drafting history out of the deliverable.
- Choose a suitable page type or template for the requested professional outcome.
- Structure the source content and produce the Markdown, HTML or JSX needed by the page.
- Prepare revisions while keeping the existing page structure and the user's requested changes intact.
Human approvals and exceptions
People keep authority over consequential and ambiguous cases.
- Decide which conversation details, evidence and claims belong in the public page.
- Approve the saved.md publishing action before the agent creates or updates a page.
- Review the rendered deliverable before sharing it with a client, colleague or public audience.
- Choose whether to update the live page, create a remix or export a fixed file.
Integration surface
Connect to the existing operating environment without pretending every tool is a system of record.
- AI workspaces
- ChatGPT, Claude, Gemini, Codex, VS Code, Claude Code, Cursor and other clients that can use the saved.md MCP connection.
- Page formats
- Markdown for direct documents, HTML for composed web pages and JSX for interactive deliverables and presentations.
- Template library
- Reusable starting points for reports, proposals, dashboards, briefs, presentations and other professional page types.
- Web and file delivery
- Public links for live reading and PDF export when the recipient or workflow requires a file.
What the engagement should deliver
Not just a demo: the operating model, working system, tests, controls, and handover needed to own it.
- Responsive SaaS web application and authenticated page dashboard
- MCP interface for agent-led page creation and updates
- Markdown, HTML and JSX publishing runtimes
- User approval boundary for publication actions
- Versioned public pages with stable share links
- Template, remix and PDF export workflows
Limitations and next phase
Start narrow enough to validate the operating model before expanding its authority or scope.
- Pages are public to anyone who has the link, so users must remove information that should remain private before publication.
- The quality of the finished page depends on the conversation context, source material and instructions available to the agent.
- Interactive JSX pages support light application behaviour, but they do not replace a full custom application when the workflow needs a dedicated backend or complex permissions.
- Connected AI tools still control how approval prompts and MCP actions appear inside their own interfaces.
- The case-study frames are official product captures from a specific release. The live website remains the source for current behaviour.
Possible next phase
Expand reusable deliverable patterns and collaborative editing while keeping publication approval and version safety visible in every agent workflow.
Your workflow will be different
Turn the blueprint into a system shaped around your tools and controls.
Bring the current steps, owners, systems, and recurring exceptions. The first conversation is for mapping the workflow and deciding whether a custom build is justified.