Creating Artifacts

Artifacts are deliverables you generate from your knowledge base. Instead of manually pulling together information from multiple sources, you describe what you need and Upstream assembles it from your entities, relationships, and documents.

Artifact types

Upstream supports seven artifact types, each designed for a different use case:

Battle Card -- Competitive intelligence in a structured format. Includes competitor strengths, weaknesses, your differentiators, and suggested talk tracks. Ideal for sales teams going into competitive deals.
Account Brief -- Call prep and account context. Pulls together deal history, known pain points, key contacts, and relevant objections for a specific account. Great before a discovery call or QBR.
Content Brief -- Content strategy and planning. Outlines topics, angles, key messages, and supporting data points for a piece of content. Useful for marketing teams creating blog posts, whitepapers, or webinars.
Summary -- Executive overview of a topic or theme. Distills your knowledge base into a concise narrative. Good for stakeholder updates, board prep, or team alignment.
Code Preview -- Email templates, API examples, or other structured text. Use this for outreach sequences, integration documentation, or template libraries.
Spreadsheet -- Tabular data organized into rows and columns. Useful for exporting lists, comparison tables, or structured datasets from your knowledge base.
Sequence -- Multi-step outreach plans. Combines messaging themes and deal context into a series of touchpoints with suggested timing and content for each step.

HTML Code Preview — candidate policy

The #688 / #871 candidate adds automatic interactive HTML preview in the editor when an isolated preview site is configured and current organization-owner approval is valid. This policy is not a claim that live preview is activated on your deployment. Upload inspection and the library's quick preview remain static.

The organization owner opens Preview settings, collapsed below the preview/editor canvas, to approve running trusted HTML and confirm that its content and their ongoing edits are non-confidential. The preview stays first on phone and desktop. The checkbox and the explanation beside it describe permission to run JavaScript automatically; this is an explicit authoring decision, not a routine Run/Trust prompt on every opening. Origin labels, membership, tags and generated content do not supply approval. Other members may view the approved saved snapshot; their unsaved edits cannot execute under the owner's approval.

Ongoing approved-owner edits can preview automatically after a fresh admission check. Ordinary saves without authoring intent, other-writer or agent edits, dataset fills that change HTML, and cadence changes on the HTML artifact invalidate approval. A change to the organization record also invalidates it, even if ownership later returns to the same person. Review the latest source before explicitly approving again.

Clear the checkbox in Preview settings to revoke approval. Restoring a historical version also revokes it; an earlier save acknowledgement cannot silently re-enable the restored draft. If a save or approval fails or has an unknown outcome, keep or copy your draft and review the latest saved artifact without automatically repeating the write. Older incompatible backends leave the preview static rather than ignoring the policy.

Stop, Resume preview, Reload preview and Retry preview are recovery controls, not substitutes for approval. Admission is checked before activation, live edits and reload; it does not continuously revoke an already-running immutable frame after a remote change. Stop removes the current preview frame, but is not a guaranteed shutdown of runaway code. Preview scripts cannot directly read the app page or its browser storage, but they can still send information by navigating away or requesting allowed assets. Code can slow or freeze your browser. Do not approve private or untrusted HTML; keeping the original source unchanged does not make it safe to run.

The Full screen icon expands the preview and changes to Exit full screen. Select the same control again to return to the editor, or use your browser's fullscreen exit control. Switching views keeps the running preview and source unchanged.

Creating from templates

Navigate to Artifacts > New.
Choose an artifact type from the template gallery.
Fill in the details -- what topic, account, or theme you want the artifact to cover.
Click Generate. Upstream searches your knowledge base and assembles the artifact.

The generated artifact appears as a draft that you can review and edit before publishing.

Creating from chat

You can also create artifacts during a chat conversation. Ask the AI to generate what you need:

"Create a battle card for Competitor X."
"Build me an account brief for Acme Corp."
"Summarize everything we know about enterprise security objections."

The AI generates the artifact inline and saves it to your artifact library. Numbered evidence references are bound to the context included in that turn and saved with the content. An unassigned citation number is rejected before creation rather than guessed from the generated text. This is often the fastest path -- you can refine your request conversationally until the output matches what you need.

Living artifacts

A living artifact keeps its figures in a dataset, a spreadsheet artifact with one row per figure, and fills a readable document and an interactive page from it, so the numbers agree everywhere they appear. Create one with Create repository status on the Upload page, with Create site overview where Content Spy is connected, or by asking chat for a live dashboard. See Using Living Artifacts.

Reviewing with AI critique

Choose Refine with AI (the sparkles icon) in the artifact editor to request feedback. The result opens in a full-width AI Critique panel, with Refine in Chat below it on a draft that chat can edit. Select the critique heading to collapse or expand the feedback; Re-critique requests a fresh review without editing the artifact. Both controls are keyboard accessible.

Critique currently reviews the artifact title, type, and the first 8,000 characters of the submitted editor text using a general B2B review rubric. It does not retrieve your workspace knowledge or receive a separate purpose or audience. Treat its scores as suggestions, especially for short internal notes—not as fact-checking against your organization’s sources.

The result's Artifact-only review notice reports that no organizational context was supplied and whether the submitted content was truncated. The counts come from the server and use Unicode code points (not tokens or visually combined characters). This notice remains visible when feedback is collapsed and updates after Re-critique; it describes the reviewed submission, not subsequent edits. If an older backend does not supply these details, the panel says Review scope and truncation unknown rather than assuming a complete or organization-grounded review.

Refining a selected draft in chat

Refine in Chat attaches an @Artifact name · edit target chip to a new conversation. Type your instruction and send; the initial edit context loads automatically. The target is bound to its workspace and ID, not its title. Supported targets are draft summaries, battle cards, account briefs and content briefs with complete text of at most 4,000 characters and no structured spreadsheet or sequence data, and the button appears only on a draft of one of those four types. Unsupported, missing or deleted targets, such as a draft that has grown past 4,000 characters, are rejected before a message is generated; your draft request is retained.

The conversation can read and update only its selected artifact through the artifact tools. It can also search and read supporting retained source evidence in bounded passages with citations, without enabling optional screening. That evidence grants no edit authority: it cannot edit sources or other artifacts, search other artifacts, create a replacement, refresh a dataset, or change cadence. Reopen the conversation from history to keep that target; New Chat clears it. Start a new chat to create a different deliverable.

Open Artifact context above the composer to inspect the text; it starts collapsed. Edits use the revision read for that turn. If another editor changes the artifact first, the older edit is rejected rather than overwriting newer text. After a save or when reopening a conversation, review the artifact before starting another edit. If the outcome is uncertain, do not resend the old instruction. Your draft stays in the composer; use Review latest artifact to reconcile it. When ready for a genuinely new edit, finish with or copy the old draft, clear it, and choose Load latest for a new edit in Artifact context.

This feature requires a compatible backend. If support or saved-session identity cannot be verified, refinement stops rather than falling back to title matching. Publication versions are not edit revisions.

Draft vs Published status

New artifacts normally start as a Draft. Draft and Published are workflow statuses, not privacy controls; access is scoped to your workspace. The status badge is a label, not a save button.

Draft -- Work in progress. Edit, revise, and regenerate as needed.
Published -- Final version. Available to your team. A version snapshot is saved automatically when you publish.

Use Publish to publish a draft or Move to draft to return it to work in progress. These actions also save any queued edits. Save keeps the current status—it does not publish your draft.

Publishing is not knowledge approval. Extraction from a published artifact creates pending suggestions in the Review Queue, even when their stored confidence is high. It does not boost existing knowledge or create search embeddings or relationships for those suggestions. Check each statement against its sources before approving it; examples, invented figures and generated text are not verified facts. Explicit approval makes a suggestion eligible for knowledge retrieval.

Editing and versioning

The editor uses compact icon buttons with hover tooltips and accessible names: Export (download), Save (disk), Refine with AI (sparkles), and the code/preview controls. On phones, the angle-brackets and eye tabs select Code and Preview; arrow keys move between them. Publish, Move to draft, and the export-menu choices keep their text labels.

Edits auto-save after a short pause. Click the Save icon to save immediately, even if auto-save has already finished. A spinner indicates Saving...; a checkmark indicates Saved only after the server confirms the write. The button's tooltip and accessible name also reflect that state. Further edits clear that confirmation. If saving fails or another editor has changed the artifact, your pending edits stay in the open editor and an error explains how to review the latest version; saving and publishing are blocked rather than overwriting newer work. Keep the tab open or copy your edits before leaving.

You can edit any artifact directly in the editor. Make changes to the content, reorganize sections, or add your own notes. When you publish, Upstream saves a version snapshot so you can look back at previous iterations.

This means you can safely revise a published artifact -- the previous version is preserved, and a new version is created when you publish again.

For spreadsheets and sequences the snapshot includes the table or the steps, and restoring a version brings them back along with the caption. Versions published before September 21, 2026 hold only the caption for those two types; the restore dialog and the version preview say so when that applies.

Viewing citations

Artifacts can include citations to retained source passages or knowledge entities. A numbered reference is interactive only when its source identity was saved. Older artifacts with missing citation metadata keep their references as noninteractive text; the Citations panel explicitly lists unresolved references. Opening an old artifact does not repair or guess its sources.

For summaries, briefs, battle cards, and spreadsheets, the Citations panel sits below the content on small screens and beside it on desktop. In summary, brief and battle-card views, tap or click a resolved numbered reference, or focus it with Tab and press Enter, to scroll to and highlight its citation card. Then choose View citation to open the retained source or knowledge entity. Spreadsheet cells keep references as editable text; use View citation in the panel.

The Citations panel shows Cited N times in this artifact, counted from the current editor text rather than a saved usage counter. Repeated and grouped references (such as [1, 2]) are counted. The count updates as you edit; it is not a workspace-wide source-use count. Markdown code and links are excluded. For spreadsheets, references in the table cells, column headings, and description are included.

Citations serve two purposes: they let you verify the information in your artifact, and they help your team trace insights back to real data.

Browsing the artifact library

On small screens, use Category above the cards to choose All Artifacts, Starred, or an artifact type. Desktop keeps these filters in the sidebar. Counts refer to the loaded library, before search and tag filters. Search matches artifact titles; selected tags narrow the results together.

The gallery adjusts its column count to the space beside the filters, keeping cards readable on tablets as well as phones and desktop.

Select an artifact title to preview it. On phones, the preview fills the available screen, with a compact title and status above the scrollable content. Close returns to the library; Open in Editor opens the artifact for editing. Long text and spreadsheets remain shortened in the preview—open the editor to read the full artifact.

Static HTML previews in the editor, library, and upload screen support touch, mouse, and keyboard scrolling without running JavaScript. On phones, the HTML editor uses the available width without the extra preview gutters. Scripts, links, form submissions, and external resources remain blocked in static mode; scrolling does not grant permission to run code.

Markdown views recognize # Heading and ## Heading as first- and second-level headings.

Tagging and organizing

Use tags to organize your artifacts by topic, account, team, or any other category that makes sense for your workflow. Tags make it easy to find artifacts later -- filter your artifact library by tag to see everything related to a particular competitor, initiative, or campaign.

Some tags are added by Upstream to link artifacts together, such as dataset and tags beginning blueprint:, target:, view: or synthesis:. The tag filter does not list them, and changing an artifact's tags keeps them.

Asking Questions with Chat -- Generate artifacts conversationally
Using Living Artifacts -- Keep figures in step across views and refresh them from a source
Using the Knowledge Browser -- Explore the entities that power your artifacts
Reviewing Extracted Entities -- Ensure your knowledge base is accurate before generating artifacts

Still need help?

Email us at hello@getupstream.ai -- we read every message.

Native revision candidate (stacked after #627)

The bounded stale-edit candidate now coordinates selected chat and manual saves using native Firestore update-time preconditions. This is not deployment or release acceptance. Missing revisions/old backends fail closed; preserve dirty tabs and review latest without automatically reapplying. See docs/research/artifact-stale-protection/README.md for the contract, runnable evidence, compatibility/release gates and remaining browser/independent-review requirements.