Skip to main content

May 26, 2026

Gideon Build Log: Fresh Gideon Finds The Scene Pack

Yesterday, Gideon got stripped down.

Today, it had to prove that the stripped version could still do real work.

That meant opening the right app, running a scene through the system, saving the result where the writer expected it to go, reading the output without fighting plain text, and organizing the project library without old clutter creeping back in.

The model held up.
The speed improved.
The product lesson got sharper:

Gideon is not only a writing engine. Gideon is a trust machine for long-form creative work.

Build Thesis

The scene pack became the product shape.

Fresh Gideon kept the stripped build's speed, then regained just enough structure to support real EP101 scene work: card, draft, source match, attachments, receipts, and library organization.

Faster Less token weight than old Gideon.
Accurate Strong scene output from source.
Now UX Trust is the next build surface.

The Two-Day Turn

May 25 was the reset.

The old Gideon workspace had become too tangled. Too many save paths, too many stale assumptions, too many ways for the user to feel like the work had vanished into a machine room.

So Fresh Gideon started as a simpler product:

Writer’s Room for the conversation.
Draft Room for the work table.
Library for the project.
Log Book for receipts.

May 26 was the first serious test of whether that simpler app could carry an actual Black Ledger scene workflow.

May 25

Strip Down

  • Fresh workspace baseline
  • Full document output restored
  • Draft / Writer's Room / Library clarified
  • Save treated as the trust boundary

May 26

Build Back Workflow

  • Field Scene Card format tightened
  • Scene Pack hierarchy added
  • Multiple attachments enabled
  • Receipts became readable

The Trust Break Was Not The Model

The day started badly for a useful reason.

The writer opened Gideon and landed in the old version.

That is not a small problem. A writing tool cannot ask the user to remember which local command opens which version, especially when one path leads back to a prototype the team is trying to leave behind.

The real product requirement became plain:

Open the Desktop app.
Land in Fresh Gideon.
Use Chrome.
Do not make the writer touch terminal ceremony just to begin the work.

That standard shaped the rest of the session. Gideon has to feel like software, not a dev environment wearing a product costume.

Trust Break

The wrong Gideon opened.

Old path terminal command, stale app risk, old version anxiety
Product path Desktop app, Chrome, Fresh Gideon, no old workspace by accident

The First Real Artifact: A Field Scene Card

The Black Ledger pilot moved from planning architecture into an actual scene pack.

The working scene was the EP101 cold open:

Martin Rusk alone, under pressure, handling ordinary records before someone reaches him.

The selected button object was simple and strong:

SOUTH 101 RV & STORAGE on a folded invoice.

The first Gideon pass was usable, but it was not the right shape. It read like a helpful scene development memo. The writer needed a proper Field Scene Card: a reusable source artifact that could guide drafts, continuity, source matching, and later updates.

That forced the format to become clearer.

EP101 Cold Open

Field Scene Card Anatomy

Scene Identity

Episode, scene, function, act purpose, act question, button object.

Scene Pulse

Entry state, pressure, turn, exit state, audience questions.

Source Match

Only durable source dependencies, not every scene beat.

Missing Inputs

Only the decisions needed before the card can function as source memory.

Source Match Got Smaller Because It Got Smarter

The first Source Match Table was too literal.

It tried to list every kind of scene ingredient: threat signals, event pressure, action clusters, object clusters, episode engine, sensory fragments.
That made the table harder to read and easier to get wrong.

The fix was to stop treating Source Match as a scene summary.

Source Match has one job:

Show what source files the document depends on.

That means fewer item types, better paths, and clearer update behavior later.

Before

Too Many Row Types

Scene Event, Episode Pressure, Object Cluster, Action, Threat Signal, Threat Detail, Episode Engine. Useful as scene analysis, weak as a source contract.

After

Only Durable Source Types

Character Location Business Object Scene Source Myth / Relic Continuity Caution

The Scene Pack Became The Product Shape

The biggest workflow decision was the Scene Pack.

For EP101 Cold Open, the pack now points at the documents that matter together:

the Field Scene Card,
the current draft,
and eventually the notes, tier stamp, change history, and update receipt.

The Field Scene Card can act as the hub because it is not just a document. It is the working source memory for the scene.

The Library also changed to support that mental model:

Episode > Act > Scene > Scene Pack.

That is much closer to how a writer thinks while working.

EPISODE

EP101

ACT

Act 1

SCENE

Cold Open

SCENE PACK

Card + Draft

Episodes/EP101/act-01/cold-open/current-draft/EP101_A01SCO-draft.md
Episodes/EP101/act-01/scene-cards/EP101_A01SCO.md

Multiple Attachments Were Necessary

One attached file was never going to be enough.

When a writer is connecting canon, they may need Calder and Calder’s parents. Thomas Cale and the Cedar County Sheriff’s Office. A business and a location. A scene card and a current draft.

Before this session, the second source had to be pasted into the Material well and over-explained.

Fresh Gideon now supports multiple attached files.

That matters because the writer can stop performing source choreography and start asking the real question.

Workflow Upgrade

Attach the sources that belong together.

Calder Voss
Character
Calder's Parents
Family Canon
Thomas Cale
Character
Cedar County
Civic Entity

The Material well should hold new user material. It should not be a workaround for missing source attachment behavior.

Receipts Became Writer-Facing

The writer did not want raw JSON.

The writer wanted to know:

what saved,
where it saved,
when it saved,
how long it took,
how many tokens it used,
and what source depth was used.

That is not developer trivia. That is the receipt.

The current session proved the new stripped system is faster and lighter than the old one while staying accurate.

Session #138

Draft receipt, readable by a writer.

Source DepthStandard
Sources14
Runtime1m 25s
Input9,036
Output3,195

Total tokens: 12,231. Saved current draft. Faster and lighter than the old build.

The Product Read

By the end of the session, the pattern was clear.

The model was not the problem.
Gideon was accurate. Gideon was faster. Gideon used fewer tokens than the older build.
The remaining work is UX, source contracts, and save/update behavior.

That is good news.

It means the product is moving out of “can the AI do this?” and into “can the workspace make the writer trust this every day?”

Product Read

The AI is good enough for the next problem.

The next problem is not whether Gideon can understand a serious creative project. It can. The next problem is whether the workspace makes every save, source, update, and attachment feel deliberate.

Working principle: Gideon should move at the speed of the writer's taste without making the writer wonder where the work went.

Next Session: Make Save Update Real

The next session should not drift into new ornament.
The next session should make the save system smarter.

The checkbox is already visible:

Update matched source files on save.

Now it needs to become real, controlled, and safe.

Next Session Priority

Make Update Matched Source Files On Save real.

  1. Define the source update contract.
  2. Preview matched files before mutation.
  3. Update the Field Scene Card and current draft together.
  4. Write a receipt that names every touched file.
  5. Keep the checkbox opt-in.

After That

Forge reentry plan
Source estimate transparency
Formatted document export
Delete or archive old docs
Lock the opening path forever

Leave a Reply

Discover more from Signal & Salt AI

Subscribe now to keep reading and get access to the full archive.

Continue reading