Skip to main content

Working with the AI chat

Glowtify is a conversational marketing assistant — most of the work happens through a chat. There are three chat surfaces today (Home, Campaign, Task), each scoped to a different context. They share the same vocabulary, the same input patterns, and the same set of tools. This article covers the common ground.

For the surface-specific behavior, see:

The input box​

Every chat has the same input affordances at the bottom:

  • Free-form text — type in natural language. Multi-line is supported (Shift+Enter for a new line).
  • File attachment — drag-and-drop a PDF, image, screenshot, or doc anywhere in the chat. The chat ingests it as additional context for the conversation.
  • @ mentions — type @ to pull in an external app via MCP. The picker shows connected apps.
  • Output mode toggle — present on the Task Chat (Text / Image / Text+Image). Controls what the next generation can touch. See Task Chat → Output mode.

Suggestion chips​

Below each reply, Glowtify surfaces suggestion chips — clickable shortcuts that pre-fill the next message. They're context-aware: the suggestions shown depend on what just happened in the conversation.

A few common patterns:

  • After generation: "Make it shorter", "Try a different angle"
  • Right after an error: "Retry", "Skip and continue"
  • When close to publication: "Schedule it", "Send for review"

A typed message works too — the chips are shortcuts, not the only path.

Target edit chips​

Specific to the Task Chat (but worth knowing across all chats): instead of asking the AI to redo the whole content, the message can be scoped to a specific part. Above the input, the Task Chat surfaces clickable chips like Edit headline, Change CTA, Replace image. Click one and the next message only touches that target.

This is the difference between "Make this shorter" (re-rolls everything) and Edit caption → "Make this shorter" (only touches the caption, leaves the image alone).

Attaching files​

Drag and drop works from anywhere on the page — the input doesn't need to be focused. Supported formats:

  • Images: JPG, PNG, WebP (used as reference style for image generation, or as visual brief)
  • PDFs: read as additional brief content (a client brief, a press release, a creative deck)
  • Docs / TXT / Markdown: ingested as text context
  • Spreadsheets (CSV, XLSX): ingested as structured data

The chat acknowledges the attachment inline and quotes the part it picked up. If a file is referenced but not attached ("here's the brief from the client"), the chat nudges to drag it in.

Brand voice — what the chat already knows about the brand​

Every reply is shaped by the workspace's Business DNA + Marketing Strategy. The chat doesn't need the brand voice re-stated in every message — it pulls the tone, vocabulary, audience, and positioning automatically.

When the chat detects a conflict between a message and the brand's guidelines (an off-brand request, or a prompt that contradicts the usual tone), it surfaces the conflict before producing. The conflict can be overridden on a per-content basis.

See Business DNA for what's stored.

External apps via MCP​

All three chats can call external apps connected through MCP — a CRM, a project tracker, an inbox, an analytics tool. Type @ to pull up the app picker, or describe the call in natural language: "Use the latest 10 leads from HubSpot to draft an email".

MCP setup is done once per workspace, not per chat. See External tools (MCP).

Memory​

Each chat keeps two layers of memory:

Within-session (short-term) — the recent message history that's sent to the model on every turn.

ChatMessages sent to the model
Home ChatLast 10 messages of today's session
Campaign ChatLast 10 messages of the campaign
Task ChatLast 5 messages of the content in progress

The full history is persisted on disk (in Firestore), but only the last few turns reach the model. Older turns can re-surface in the prompt via short-term memory recall — best-effort, lossy — and shouldn't be relied on for critical context.

Across-session (long-term) — Brand Memory — durable facts the chat learns about the brand while you converse. Stored separately from the message history, shared across all chats (Home, Campaign, Task), and persists over time (with mild decay). Brand Memory is automatically extracted from user messages — the assistant's replies are not stored as facts.

Brand Memory is not the same thing as the Business DNA:

Business DNABrand Memory
TypeStructured profile, editable by the teamImplicit facts inferred from conversations
Where it's controlledOnboarding, Brand DNA reviewLives behind the scenes (not directly editable today)
BehaviorDeterministic — always reads the sameDecays over time, weights newer facts higher
ScopePer storePer store, shared across all chats

When the recent conversation conflicts with Brand Memory, Glowtify follows the recent conversation — the most recent direction wins.

Caveats:

  • The Task Chat truncates to the last 5 messages. Long iteration threads on a single content piece can lose early context — restate constraints when the chat seems to forget.
  • Memory recall has hard timeouts (a few seconds). If the recall service is slow, the chat may answer without the Brand Memory layer, which can show up as "the chat forgot something it usually knows".
  • Personal info (emails, phone numbers, API keys) is redacted before any memory is stored.
  • Anonymous / unauthenticated stores have no memory at all.
  • Deleting a content piece or a campaign wipes that conversation's session memory but leaves the long-term Brand Memory intact.

What the chat sees​

It helps to know what context each chat has access to:

ContextHome ChatCampaign ChatTask Chat
Business DNA✓✓✓
Marketing Strategy✓✓✓
MCP-connected apps✓✓✓
The current campaign's brief—✓✓
The current content's body, channel, history—partial✓
Other contents in the campaign—✓—
The calendar / publication queue✓✓partial

A chat will never ask for context it already has. If it does, that's a bug — contact support.

Conversational style​

Some things to know about how Glowtify talks back:

  • Replies are short by default. Ask for a longer explanation and one will come.
  • The chat asks before doing something with side effects. Generating content, scheduling a post, sending for review — these surface a confirmation step.
  • The chat takes the last message most seriously. A long argument thread carries less weight than a recent, concise restatement.
  • The chat is bilingual — write in French and get French replies. Write in English and get English. Switch mid-thread and it follows.

Good habits​

A few habits that make the chat work better:

  • Be specific about what needs to change, not what's wrong. "The headline is too long" is weaker than "Shorten the headline to under 8 words".
  • Reference the part being discussed — "the second bullet", "the CTA", "the image in the top-right". Better yet: use a target edit chip.
  • State the constraint, not just the goal — "Make it punchier, max 100 characters" gets closer than "Make it punchier".
  • When the chat drifts, restart the brief in one paragraph instead of arguing. A long thread accumulates contradictions; one clean paragraph resets the context.

Common errors​

A few patterns and what to try first:

SymptomFirst thing to try
Generation completes but the output didn't changeCheck the output mode — wrong mode is the #1 cause
AI claims it edited but the change isn't visibleRefresh the page. Known issue
Reference image not honoredRephrase the prompt to mention the reference explicitly
"That didn't go through" errorRetry once. If it persists, contact support
Chat refuses the requestThe chat is likely hitting a guardrail (brand voice, content policy). The reply explains which.