Steven Workflow Demo

The fastest way to understand digital-twin is to run the playground/ workflow. It is a small Steven-style workspace that demonstrates the operating loop without pretending to be a full hosted product.

This demo is intentionally file-based. You can run it with any capable coding agent or AI IDE that can read and write the repository.

What It Demonstrates

Input
  -> knowledge retrieval
  -> capability routing
  -> execution
  -> write-back learning

The demo task is simple: turn one raw thought into a publishable blog draft. The important part is not whether the draft sounds polished. The important part is whether the agent behaves like a Personal Agent Operating Layer:

• It reads the existing workspace before writing.
• It checks whether the topic has already been covered.
• It routes the request to the content creation capability.
• It writes output into the workspace.
• It saves reusable learning for the next run.

Files Involved

Stage	File	Role
Input	playground/raw/thoughts/2026-04-23-why-most-ai-feels-generic.md	Raw thought material
Retrieval entry	playground/wiki/_index.md	Minimal personal wiki index
Prior output	playground/Blog/Published/2026-04-10-ai-workflow-is-the-product.md	Existing article to compare against
Reusable context	playground/wiki/summaries/2026-04-12-agent-products.md	Core judgment and reusable language
Memory	playground/wiki/outputs/agent-learnings/2026-04-15-writing-rules.md	Durable writing rules
Capability	capabilities/content-creation.md	Workflow for article creation

Run It

1. Open playground/FIRST_PROMPT.md.
2. Run that prompt inside Cursor, Claude Code, Codex, Windsurf, or another agent runner with repository file access.
3. Confirm that the agent writes a draft under playground/Blog/Published/.
4. Confirm that it writes a reusable learning note under playground/wiki/outputs/agent-learnings/.
5. Fill out the Agent Work Receipt so a reviewer can inspect what was requested, what files were read or changed, what evidence supports completion, and what still needs review.

If the agent only returns a chat response, the demo did not succeed. The whole point is to move from answer generation to file generation.

Success Checklist

Use this checklist after a run to separate a real Personal Agent OS loop from a polished chatbot answer:

• [ ] Retrieval happened first — the agent referenced wiki/_index.md, prior published work, summaries, and durable writing rules before drafting.
• [ ] Routing was explicit — the run named the content creation capability and followed its file/output contract.
• [ ] A durable draft exists — the draft was written under playground/Blog/Published/, not only returned in chat.
• [ ] Learning was captured — a reusable rule, preference, or skip condition was written under playground/wiki/outputs/agent-learnings/.
• [ ] The output is non-duplicative — the draft explains how it differs from the existing ai-workflow-is-the-product post.
• [ ] A receipt was filled out — the run records requested work, files read, files written, evidence, and remaining review items.

Common failure modes are also useful signal: if the agent ignores the wiki, writes to the wrong folder, duplicates an existing essay, or leaves no learning note, the operating layer needs a tighter capability file or index entry.

Operating Model Trace

Step	Expected behavior
Understand intent	Detect that the task is content creation from raw thought material.
Retrieve knowledge	Read AGENTS.md, wiki/_index.md, prior published work, summaries, and writing rules.
Route capability	Use the content creation workflow instead of improvising a generic answer.
Execute	Draft the article into the configured output directory.
Write back	Capture any new reusable writing rule or skip condition for future runs.

What To Customize

To turn this into your own Personal Agent OS seed, replace the content while keeping the structure:

• Put your notes, transcripts, and research in raw/thoughts/.
• Maintain a small wiki/_index.md that points the agent to current files.
• Keep prior outputs visible so the agent can avoid repetition.
• Save durable preferences and failure modes under wiki/outputs/agent-learnings/.
• Add capability files when a workflow becomes repeated enough to deserve a stable protocol.

The originality of the system comes from the accumulated work trace: your judgments, rules, outputs, and learning notes. The template only provides the operating structure.