Where did we work that out?
Find one detail, load the relevant context, then follow connections across the archive.
Studio / Engine / Connections / Memory / Workbench
The work
inside the work.
Try another image direction without losing the first. Inspect a tool boundary. Recover the source of a half-remembered decision. The details are where the work becomes useful.
Thirty source-backed stories, with small examples you can inspect and evidence you can open. Choose a question, or filter the catalogue.
Three ways in
Pick a useful question.
Why should this fact stay?
Resolve a name, inspect its source and keep a reasoned history of corrections.
How does this work carry across sessions?
Share the setup, give browser work an owner and return evidence for review.
The feature atlas
Specific problems.
Named features.
I build the smaller pieces that make a workspace useful. Explore image revisions, agent streams, tool boundaries, memory and the workbench around my coding agents.
30 feature stories
Dates mark a researched episode or release anchor.
No features match that combination. Try a broader topic or clear the filters.
- StudioReference to chat
Take the image straight into the conversation.
A visual reference becomes cumbersome when it must be downloaded and reattached before asking for a revision.
- StudioBranching workbench
Keep the good answer. Try another direction.
Changing an earlier instruction can erase the path that produced a useful answer.
- Workbenchmetaclaude
Give reusable agent setup a versioned home.
Agent setup spreads across instructions and configuration files. Manual copies drift, and copying a whole runtime directory includes state that should stay local.
- WorkbenchReusable skills
Package a repeated task with a clear contract.
Explaining a familiar procedure from scratch makes it easy to omit a step. A bare tool name also says little about when or how it should be used.
- WorkbenchBrowser evidence
Keep a browser result connected to its task.
A screenshot without a task, action or page identity is difficult to interpret later. Browser exploration can otherwise become detached from the agent’s report.
- EngineSee what the agent is doing
Read the work around the answer.
A streaming answer can look stalled or opaque when thinking and tool work are hidden inside a plain text bubble.
- ConnectionsA voice conversation with visible state
Make a spoken conversation visible.
Hands-free conversation needs a clear path through listening, transcription, response and spoken playback, rather than a hidden recording control.
- MemoryKnowledge graph
Follow an idea across conversations.
A summary helps with one conversation. An archive of separate summaries still makes it hard to see how a project, a decision and a later change fit together.
- MemoryGraph exploration
Make a large graph readable.
More nodes bring more crossing lines. Showing every label at once can make a graph harder to read and slower to explore.
- MemoryPersonal relevance
Keep the facts that distinguish the work.
A personal graph can fill up with facts a reference book already knows. Those records make specific decisions and durable context harder to find.
- EngineThree engines, one chat contract
Change the runtime. Keep a readable chat.
Supporting several agent runtimes can fragment the chat UI and force client code to understand each engine's streaming protocol.
- MemoryFindThat
Recover the detail you half remember.
A remembered incident is often a poor search phrase. Broad results also leave the reader to decide which session actually contains the useful detail.
- MemoryMaintained profile
Keep a working brief with a restore path.
Repeating stable working preferences takes time. An automatically maintained brief needs visible history so its changes can be checked.
- MemoryGrounded analysis
Resolve the names. Keep the source close.
A model can save a neatly structured statement and lose a qualification. Names and source links help inspect the result; they do not make every claim correct.
- MemoryContext across engines
Carry an approved brief into a fresh session.
A working brief can become tied to one provider. Switching engines then means explaining the same stable context again.
- MemorySame-Page
Bring useful context into the next conversation.
Replaying a transcript adds noise. It can also make unfinished work from an old session look like a new instruction.
- StudioGive a creative library a home
Move the project. Keep its contents together.
A growing gallery needs a stable way to group projects, scenes and panels without flattening every asset into one long list.
- MemoryCuration history
Give a correction a reason and a history.
Cleanup is hard to assess when a fact disappears or changes without leaving a reason. A mistaken correction also needs a way back.
- WorkbenchHelm
Review delegated work across agent runtimes.
A delegation practice tied to one assistant breaks when the lead changes runtime. Child reports also need a review step before their suggestions become guidance.
- Workbenchmetacodex
Share Codex setup while keeping local settings.
A machine may need its own settings while still sharing the same instructions and hooks. A blanket configuration copy can overwrite those differences or carry credentials with it.
- WorkbenchWorkbench sync
Update the shared setup without copying local state.
A configuration update should reach enrolled machines while preserving their credentials, runtime history and valid local differences.
- EngineRecover the answer after a disconnect
Close the window. Keep the result.
A browser or proxy connection can disappear while a long-running tool turn continues on the engine.
- WorkbenchModel policy guard
Check what an agent launch would use.
A valid model and effort on the request can still be overridden by a configured default role.
- EngineRedirect a turn without throwing it away
Change the instruction. Keep the work.
A useful result may already be on its way when the answer needs a different format. Interrupting the request can discard that progress.
- WorkbenchBrowser handoff
Send the page to the session that owns it.
Several sessions can share one browser. Without ownership, a page opened for one task can be mistaken for another session’s work.
- ConnectionsA tool gateway with clear boundaries
Check who, what and which tool separately.
Tool names alone do not explain who may call a resource, which tools a profile exposes, or which OAuth scope permits the call.
- ConnectionsAsk the person who can answer
Bring a waiting decision back to the work.
An agent may need a human decision when the person is away from the pane where the work is running.
- WorkbenchFlow screenshots PDF
Make a browser walkthrough someone can review.
A folder of screenshots loses the order of a user flow. Different screen sizes are also difficult to compare without consistent labels and captions.
- EngineChoose the tools an agent may use
Make tool choices an enforced boundary.
A profile that names only allowed MCP tools leaves native shell, browser or other runtime tools outside the visible choice.
- WorkbenchObserver screenshots
Give the human a view of the browser action.
A remote supervisor needs to see whether a browser interaction reached the expected screen. Adding every image to model input also increases the payload.