May 6th, 2026
GIDEON Build Log: Active Scene Memory
We are building Gideon in public because the product is teaching us what it wants to become.
May 6 was the foundation day.
The goal was not simply to finish a scene, though useful scene material came out of the session. The real goal was to test whether Gideon could support the movement from idea to Scene Work Card to Forge tiers to draftable material without losing the writer’s intent.
The first major product law became clear:
Gideon is not valuable because it replaces the writer.
It is valuable because it holds context, asks sharper questions, preserves decisions, and helps the writer move from idea to draft with more clarity.
That is the difference between a writing tool and a prompt machine.
What We Proved
The Scene Work Card workflow is working.
For EP203_A01S01 – End of the Tour, Gideon produced a strong Field Scene Card with the pieces a writer actually needs:
– scene identity
– scene pulse
– source match table
– location
– visual iconography
– dialogue snapshots
– emotional role map
– myth, relic, and fracture tags
– threat pressure
– next-scene link
– missing inputs
– recommended next tier
A scene cannot be treated like a passing chat prompt.
The output was clear, structured, and copyable. More importantly, it gave the scene a working identity.
That matters because a scene cannot be treated like a passing chat prompt. If the writer is going to return to it, revise it, validate it, and carry it through multiple passes, the scene needs memory.
The Forge Started To Behave
We tested the movement:
Scene Work Card
-> T1 Thread the Bones
-> writer choices
-> T2 Salt the Blood
-> draftable material
T1 worked because it did not try to write the scene too early. It gave beat structure, actable character jobs, open choices, and pressure points.
T2 worked better once writer notes were integrated. It expanded emotional pressure, atmosphere, and scene texture while still flagging canon-sensitive material.
That is the behavior we want.
Gideon should not flatten every request into “here is the finished scene.” It should help the writer understand what stage the material is in, what choices are still open, and what kind of pressure the next pass should apply.
GIDEON Needed To Remember The Scene
The biggest failure of the day was not creative.
It was state.
At one point, a full-scene request drifted from End of the Tour into a different EP203 thread. Gideon had enough story intelligence to help, but the app had not yet given it a stable active scene contract.
So we built Active Scene Memory.
Active Scene now carries:
– scene identity
– scene card
– latest T1
– latest T2
– latest draft output
– writer notes
– active scene draft, when available
That changed the feel of the app immediately.
The writer no longer has to manually rebuild the context every time. Gideon can stay anchored to the scene being worked on.
Biggest failure of the day: state.
The Active Scene Bar Became A Control Surface
The Active Scene Bar also became more than decoration.
It now includes:
– Draft / Rewrite mode toggle
– tier buttons from Chart through T6
– selected tier highlighting
– View Card
– Clear
The important part is not the buttons.
The important part is that Gideon started carrying the workflow burden. A writer should not have to remember prompt wrappers or internal phrasing just to make the system behave.
That is what Signal & Salt AI workspaces are for.
Cleaner Chat, Better Tool
Opening a scene used to paste the full scene card into the Writing Session composer.
That worked technically, but it made the app feel like visible prompt assembly. The writer had to watch the machinery.
Now, opening a scene sets Active Scene silently and keeps the composer clean.
That is a quiet product improvement, but an important one. Gideon starts to feel less like a chat window and more like a writing-room app.
The Product Model Started To Separate
This session clarified three major objects.
Scene Card
is the working source object.
It tells Gideon what the scene is, what sources matter, who is present, what pressure is active, where the scene exits, and what flags are live.
Forge Trail
is the process history.
Each tier should eventually leave a visible stamp:
– Chart: direction formed
– T1: structure pass complete
– T2: expansion pass complete
– T3: canon validation complete or flags open
– T4: character consequence complete
– T5: voice and format lock complete
– T6: final seal or not sealed
Decisions
are the governance layer.
This is where canon updates, major changes, rule locks, session recaps, and “we changed the scene because X” should live.
That framework matters because it turns Gideon from an AI helper into a system with memory, process, and accountability.
What Broke
The failures were useful because they showed us exactly where Gideon needed more spine.
A Scene Work Card request returned a Matrix Walk. That was a routing problem. Explicit scene-card requests now route to the scene-card path and are blocked from returning Forge-style output.
Scene ID confusion showed up around EP203_A01S01 and EP203_A01S02. Gideon flagged the mismatch, and the scene was corrected in the repo.
Gideon did not show EP203 as its own episode group until repo-native files existed. Once the files were added, Gideon indexed the episode correctly.
Gideon re-asked questions that had already been answered. The prompt contract now needs a stronger “Captain Notes Integrated” pattern.
Raw HTML leaked into output. Gideon now asks for clean markdown only and strips common HTML before rendering.
The Save Trail is still conceptual. The tiers produce useful outputs, but tier stamps are not yet durable scene state.
That last point became the next major milestone.
Product Truth
The most important insight from May 6 was simple:
Gideon asks better questions.
The writer answers with taste and intent.
Gideon turns those answers into stronger structure.
The writer writes with more clarity.
Gideon expands, stress-tests, validates, and seals.
That is the value.
Not replacement.
Pressure. Structure. Continuity.
What Comes Next
The next major direction is to make the Save Trail real.
Tier stamps, linked decisions, and scene status need to become durable parts of the scene packet instead of loose session behavior.
That is how Gideon becomes safer for real work.
Not more chaos.
More spine.
