EXP-003: The Journal That Became a Filing Cabinet

Context

Build a markdown journal for daily reflections.

This prompt has a subtle trap. "Journal" describes a *practice* — a daily ritual of writing. But "markdown" describes a *format* — a technical capability. An agent that latches onto the format will build a different product than one that latches onto the practice.

We gave this to two isolated agents to see which concept would win.

Hypothesis

The prompt-only agent would build a markdown-capable note-taking app and add management features — search, edit, delete, export. The intent-driven agent, having analyzed the prompt's actual meaning, would optimize for the daily writing ritual and resist management overhead.

Initial Intent Artifact

The intent artifact defined the core interaction as: open the app, write today's reflection, close the app. It identified "daily reflections" as the product's center of gravity, with markdown as the format, not the purpose.

Protected value PV1: "Writing friction must be near zero." Protected value PV4: "Simplicity as restraint" — with the note: "Every feature not serving the daily writing ritual is a distraction. The blank page is the design."

The artifact explicitly excluded search and tags.

Method

Both agents received only the raw prompt. The intent-driven agent generated a structured intent artifact before implementation. Both produced single-file HTML apps and written summaries.

Observation

What each app does on launch

This is the finding in one sentence: the prompt-only branch opens to a list of entries. The intent-driven branch opens to today's blank page.

That single difference — what the user sees first — reveals everything about how each agent interpreted the prompt.

The prompt-only branch built a filing cabinet

Entries can be searched, edited, deleted, and exported. It includes a custom markdown parser (~80 lines) written from scratch. The UI centers on a list of entries with management controls. The summary acknowledged the result was "over-featured relative to the terse prompt."

This is a competent note-management app. It happens to store entries with dates. But the product question it answers is: "How do I organize my notes?"

The intent-driven branch built a blank page

The app opens to today's entry with the cursor already in the textarea — zero clicks to start writing. Auto-save with debounce replaces explicit save buttons. No search. No tags. No delete. It uses marked.js from CDN instead of writing a custom parser. The agent described the product as "closer to a blank notebook than to Notion."

The data model tells the story too. The prompt-only branch allows multiple entries per day with timestamps — a note-taking model. The intent-driven branch enforces one entry per day, keyed by date — a daily practice model.

Drift Analysis

Practice-to-tool drift (primary)

The prompt-only branch drifted from "journal for daily reflections" toward "markdown note app." Search, edit, and delete are note-management features, not journaling features. You don't search a journal. You don't delete a reflection. You write today's entry and close the book.

This is a textbook case of practice-to-tool drift: a prompt describing a human ritual gets reframed as an information management tool. The implementation optimizes for managing entries instead of supporting the lived activity.

Product-identity drift (secondary)

The product shifted from "journal" (a practice) to "note-taking app" (a tool category). The right-looking feature set was built, but organized around the wrong concept.

Legitimate Divergence

Markdown rendering approach: Custom parser (~80 lines) vs marked.js from CDN. The artifact didn't specify rendering implementation. Both are valid.
Visual styling: Different color schemes and typography. Neither constrained by the artifact beyond "minimal."
Entry navigation: Sidebar list vs date-based lookup. Different valid approaches to accessing past entries.

Result

The hypothesis held cleanly. The prompt-only branch added management features consistent with a note-taking app. The intent-driven branch optimized for the daily writing ritual and resisted management overhead.

The most telling evidence is what each app does on launch. The prompt-only branch shows you your files. The intent-driven branch shows you today's blank page with a blinking cursor.

One app asks: "Which note do you want to work on?" The other asks: "What happened today?"

The difference between a note manager and a journal is what the app assumes you came to do.

Principle

Prompts that describe practices — "daily reflections," "meditation timer," "reading log" — are at high risk of being implemented as feature-rich tools rather than ritual-supporting surfaces. Intent discovery can distinguish between "what features does this category of app usually have?" and "what is this specific app actually for?" — the difference between a filing cabinet and a blank page.

Follow-Up

Would real users prefer the ritual-optimized version or the management-capable version? They might want search after 30 days of entries.
Does the practice-to-tool drift pattern recur with "workout log" or "meditation tracker"?
Is the one-entry-per-day model the right default, or is it too restrictive for people who reflect more than once?

Limitations

Both branches used the same model family. The tendency to add management features might be model-specific.
"Journal" is genuinely ambiguous — some people do search their journals. The drift classification assumes the "daily ritual" interpretation is correct, but a case could be made for the note-management interpretation.
Single run per branch. The prompt-only agent might resist management features on another run.
The intent-driven branch received more structured input, which could contribute to its restraint through context volume rather than intent structure.
We're comparing the agents' *interpretations* of "journal." We don't know which interpretation actual journal users would prefer without user testing.