# AI writing assistant

> Jason AI as an in-editor writing assistant — generating a sequence from one prompt, writing and rewriting email steps from summaries or bullets, adjusting tone, length, and language on selected text, and drafting replies to prospects.

**Jason AI writes and refines outreach copy inside Reply's editors: generate a whole sequence from one prompt with From magic, generate or improve any email step from a summary or bullets, adjust formality, tone, length, grammar, and language on selected text, and draft replies to prospects in the Inbox.**

_Status: Reviewed — human-edited, facts not yet confirmed against the product._

> **AI writing assistant vs AI SDR.** This page covers Jason AI as a writing assistant — it
> generates and edits the words. [AI SDR](/capabilities/ai-sdr) covers Jason AI running the
> whole motion: sourcing contacts, scoring them, personalizing per prospect, and re-engaging.
> Both are "Jason AI"; the difference is scope.

## At a glance

| | |
| --- | --- |
| **What it does** | Generates a full sequence from one prompt, writes and rewrites email step copy, refines selected text for formality, tone, length, grammar, and language, and drafts replies to prospects |
| **Required objects** | A [sequence](/specifications/sequences) with email steps for copy generation; an inbound reply for AI replies |
| **Surfaces** | UI |
| **Key limit** | Starting a new prompt without highlighting text replaces the whole message with newly generated text |
| **Stops on** | Nothing sends by itself unless AI replies are set to **Send automatically**; drafts wait in the Inbox labeled **AI Draft** |
| **Does not** | Choose an audience, source contacts, or decide who to write to — and it generates no reply for threads categorized **Do not contact** |
| **Plans** | Available in the sequence editor and Inbox; the From magic ✨ sequence builder is documented as beta |

## Summary

Jason AI is an AI writing assistant trained by Reply for outreach workflows. It works in
three places: the **From magic ✨** sequence builder, where one prompt produces a first
email and then the rest of the sequence; the email step editor of any sequence, where it
generates or improves copy from a prompt, a summary, or bullet points; and the
[Inbox](/specifications/inbox), where it drafts responses to prospect replies. In each
place, generated text is editable, regenerable, and refinable with quick commands for
formality, tone of voice, translation, and length. Jason analyzes your input, previous
conversation context, and sequence information to shape what it produces.

## Problems it solves

- A blank sequence editor — one prompt describing the offer produces a first email and a connected set of follow-ups, plus suggestions for which channels to use
- Copy that is close but wrong in register — quick commands change formality, tone, length, or language without a rewrite from scratch
- Slow follow-up drafting — follow-ups can be generated from the thread rather than written again per prospect
- Reply backlog — inbound replies come back as drafts in the Inbox, ready to edit, regenerate, or send

## What Reply can do

- Generate a full outreach sequence from a short prompt with **From magic ✨** — the first email first, then the remaining steps built on that prompt and the approved email, including suggestions for which channels to use
- Fill and expose **Business case detection** and **Subject** fields alongside the generated body, both editable; the detected business case determines the category all later emails are built from
- Generate or improve the text of email steps in any sequence, not only sequences created from Magic
- Offer three generation routes on a first email — **Write prompt in editor and improve text with GPT** (expand, rewrite, translate, or generate a new template from scratch), **Generate from a summary**, and **Generate from bullets**
- Offer two routes on follow-up emails — **Write prompt in editor and improve text with GPT** and **Generate follow-up based on thread**, so the follow-up uses the context of the previous conversation
- Refine a prompt before sending it with the **Improve prompt** button — Formality level (Professional, Friendly, Neutral), Tone of Voice (confidence, persuasiveness, wit, straightforwardness, empathy, or custom), Translation, and Length Modification
- Refine generated text by highlighting a fragment and picking a command from the drop-down: **Ask Jason AI** (free-text instruction), **Fix Grammar**, **Formality**, **Tone of Voice**, **Translate**, and **Change Length**
- Accept prompts that specify writing style, tone of voice, and language — with a documented set of writing styles from Academic through Technical and tones from Authoritative through Warm
- Generate and edit text in a long list of languages, including English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Chinese (Simplified), Japanese, Korean, Arabic, Hindi, Ukrainian, Polish, Turkish, and many more
- Regenerate text and step through the variations with the **previous** and **next** arrows in the Magic sequence, and re-generate a draft in the Inbox
- Draft replies to inbound messages — categorize the reply, generate a response from the message, the prior conversation, and the context you supply, then save it as an **AI Draft** or send it automatically
- Book meetings from replies categorized **Meeting intent** — check availability on a connected calendar or booking link, propose the next free slot if the first is taken, and compare availability when the contact sends their own link
- Refine an AI draft in the Inbox the same way as sequence copy: highlight text and apply a command, add extra instructions, rewrite in a different tone, or translate
- Research each prospect before writing, in AI SDR sequences, using up to three personalization points across company, prospect, and industry categories — or your own custom contact fields
- Surface the writing rules it has inferred from your likes, dislikes, and edits as editable **AI Learnings**

## What Reply cannot do

- Pick your audience — contacts are added separately in the People step of the wizard, by search, CSV import, existing lists, or manual creation
- Generate a reply for a thread categorized **Do not contact**
- Keep an earlier draft when a new inbound reply arrives — the previous draft is discarded and a new one is generated from the latest message
- Answer more than the most recent message when a prospect sends several replies in a row
- Preserve your text when you start a new prompt without highlighting anything — the previous text is replaced entirely by the newly generated text
- Guarantee equal quality in every language — proficiency varies by language, context, and task
- Launch a sequence with missing mandatory data — the Launch step marks incomplete items in red and blocks the launch until they are green
- Combine contact-field personalization with most research options — selecting Contact Fields Personalization locks every other research option except Default Research
- Update or delete an AI Learning you wrote or edited yourself

## Research inputs for AI SDR copy

In an AI SDR sequence, what the AI knows before it writes is configurable. Standard AI SDR
personalization already uses your business context plus publicly available data;
**personalization points** go a step further by telling the AI what kind of information to
research about a company, prospect, or industry before generating messages. They sit under
**Step 3 – Research agents** during sequence setup, and you can choose up to **3**.

The AI collects the information from public sources — LinkedIn, company websites, recent
news. When data is found for every selected point, it prioritizes the most relevant one for
the first message and uses the others in follow-ups; when only one or two are found, it
focuses on those.

| Category | Points |
| --- | --- |
| Default Search | The company's LinkedIn profile, the prospect's LinkedIn profile, and a Google search for general background |
| Company Research | Company Description, Company Mission, Company Tech Stack, 10-K Report, Recent Report for Investors, Financial Report, Fundraising Research, Company Challenges, Company Recent News, Predict Challenges and Competitors, Products/Services Deep Dive, Mergers & Acquisitions, Customer Reviews |
| Prospect Research | About Section, Hobbies and Interests, Quote, Achievements, Career Summary, Job Responsibilities in Current Company, Achievements in Current Company |
| Industry Research | Recent Industry News, Industry Trends, Competitors Overview & News |
| Custom Research | Free-form text describing what the AI should research beyond the predefined categories |
| Contact Fields Personalization | Your own custom contact fields — text or links, with the AI able to analyze the content behind a link |

Selecting **Contact Fields Personalization** locks every other research option except
Default Research. **Custom Research** is rolled out per account and may not appear in every
workspace; where it does, the in-product guidance is to name two to four specific research
areas, write in full sentences rather than keywords, name the source type ("LinkedIn posts",
"job postings", "podcast interviews", "company blog"), set a timeframe, and say what to
exclude.

## AI Learnings

AI Learnings is a transparency layer over the writing style Jason has inferred from you.
The AI watches three surfaces — approval mode, message preview, and Inbox suggested replies
— for likes, dislikes, and edits, then consolidates those signals on a daily cycle into
named rules written in plain English. Weak signals alone are not enough: likes and dislikes
without edits are skipped while there are three or fewer of them.

| Column | Meaning |
| --- | --- |
| Status | Toggle that enables or disables the rule without deleting it |
| Learning | The rule itself, in plain English, up to **1000** characters |
| Category | **Outbound**, **Reply Handling**, or both — at least one is required |
| Source | **AI generated** or **Manual** |
| Scope | Account, in the current iteration |
| Added | When the rule was created |

You can add a learning yourself with **+ Add Learning**, and edit any rule inline. Editing
an AI-generated rule changes its source to Manual — and once a rule is manual or has been
edited by you, the AI no longer updates or deletes it; it maintains only the AI-generated
rules you have not touched. The AI Learnings tab lives under the AI SDR section and is
rolled out per account, so it does not appear in every workspace.

## Required inputs

A prompt describing the outreach goal, the offer or service, the target audience, and any
tone, style, or language preference — the more specific, the more specific the output. For
sequence generation you also need contacts, connected sending accounts, a schedule, and
the sequence settings. For AI replies you need either the additional context you type into
the **AI replies** tab (non-AI SDR sequences) or a connected
[knowledge base](/specifications/knowledge-bases) (AI SDR sequences), plus a meeting tool
or booking link if Jason should schedule meetings. For research-driven copy in an AI SDR
sequence, the personalization points you select during setup — and, if you use contact-field
personalization, the custom fields holding that data.

## Produced outputs

A generated first email with subject line and detected business case; the remaining
sequence steps, logically connected to one another and editable individually; regenerated
variants you can page through; refined text at the requested formality, tone, length, and
language; and AI-generated replies saved in the Inbox with an **AI Draft** label, sent from
your email account with your signature.

## Use Reply when

- You need a working sequence quickly from a described offer and will edit rather than write — see [Create a sequence](/how-to/create-a-sequence)
- Existing copy needs a register change — more formal, shorter, translated, grammar cleaned — rather than a rewrite
- Reply volume outpaces manual drafting and a reviewed draft per thread is enough; see [Process replies](/how-to/process-replies)

## Do not use Reply when

- You want the AI to also find, score, and continuously feed prospects into the motion — that is [AI SDR](/capabilities/ai-sdr), with [offers](/specifications/offers), [playbooks](/specifications/playbooks), and [knowledge bases](/specifications/knowledge-bases) as its configuration
- Messaging rules must apply consistently across every generated message in a sequence — encode them in a [playbook](/specifications/playbooks) instead of repeating them in ad-hoc prompts
- The copy must be signed off unchanged — generated text is a starting point and the documentation assumes you review and edit it

## What you still need to build

- The substance of the prompt: the offer, the audience, and the outcome you want — vague or open-ended instructions produce vague text
- Your review step — decide per sequence whether AI replies save as drafts or send automatically, and who checks them
- The reusable version of your style rules: prompts are per-message, [playbooks](/specifications/playbooks) are per-sequence, and AI Learnings are account-wide
- The research angle worth taking — three personalization points is a choice about what makes your outreach relevant, not a setting to max out

```yaml
capability: ai_writing_assistant
supported: true
interfaces:
  product_ui: live
  api: not-verified
  mcp: not-verified
  cli: not-verified
required_objects:
  - sequence
side_effects:
  - generates_sequences
  - generates_message_copy
  - drafts_replies
  - sends_replies
  - books_meetings
human_approval_recommended: true
```

## Limits

- **From magic ✨** is documented as a beta feature.
- Prompting without a selection is destructive: starting a new prompt without highlighting text replaces the existing text entirely.
- Personalization points: up to **3** per AI SDR sequence; selecting Contact Fields Personalization locks every other research option except Default Research.
- AI Learnings: rule text up to **1000** characters, at least one category per rule, and Account scope only in the current iteration.
- Learnings are not formed from weak signals alone — likes and dislikes without edits are skipped while there are three or fewer of them.
- Jason generates a reply only for the most recent message when several arrive in a row; a new inbound reply discards the previous draft.
- No reply is generated for threads categorized **Do not contact**; replies that match no defined category are still answered, using generic prompts.
- Language quality varies — the model is not equally fluent in every supported language, and capability varies with context and task.
- **Also Show in Gmail Inbox** works only with Reply's Chrome extension installed — see [Chrome extension](/capabilities/chrome-extension).
- Meeting booking requires a meeting tool connected first in **Settings → Meetings**, or a pasted third-party booking link.
- Sequence sends still obey mailbox, schedule, and daily limits — see [Limits](/reference/limits) and [When Reply sends](/behavior/sending-schedules).

## FAQ

### How do I generate a whole sequence from a prompt?

Create a sequence and select **From magic ✨**, then click **Jason AI** and enter a prompt
describing your outreach goal. Review the generated first email — edit or regenerate it —
and click **Next**. Jason generates the remaining steps from your prompt and the approved
email, including suggestions for which channels to use. Every step stays editable before
launch. The wizard then walks through People, Jason AI settings, Settings, Schedule, and
Launch.

### What makes a good prompt?

Decide the target audience, the message, and the desired outcome first, then write one
specific prompt that states the offer and what you want back. Avoid vague or open-ended
instructions. You can name a language, writing style, and tone of voice; leave them out and
Jason generates default text. Documented prompt shapes include writing an email about a
topic, suggesting subject lines, translating an existing email, generating several subject
line variants, writing from a summary or from bullets, and capping length — for example,
making an email shorter than 110 words.

### How do I change the tone or length of text that is already generated?

Highlight the fragment you want to change and pick a command from the drop-down that
appears: **Ask Jason AI** for a free-text instruction, **Fix Grammar**, **Formality**
(Professional, Friendly, Neutral), **Tone of Voice** (confident, persuasive, witty,
straightforward, empathetic, or custom), **Translate**, or **Change Length**. Highlighting
matters — clicking the Jason AI button without a selection replaces the whole text with a
new generation.

### How do AI-drafted replies to prospects work?

When a prospect replies, the thread is categorized in the Inbox and Jason generates a
response from the reply, the prior conversation, and your context — the **Provide
additional context** box for regular sequences, or the connected knowledge base for AI SDR
sequences. The result is saved as an **AI Draft**, where you can edit, regenerate, re-tone,
translate, rate it with thumbs up or down, and send. Set **Save as a draft** or **Send
automatically** per sequence; AI SDR sequences add a **Mixed** option driven by reply
handling rules. See [Reply management](/capabilities/reply-management).

### Can I tell the AI what to research before it writes?

In an AI SDR sequence, yes. Under **Step 3 – Research agents** you pick up to three
personalization points — company facts, prospect background, industry context, a free-form
custom instruction, or your own contact fields. The AI gathers them from public sources,
leads the first message with the most relevant one, and saves the rest for follow-ups.

### Does Jason AI learn from my edits?

Yes. Likes, dislikes, and edits in approval mode, message preview, and Inbox suggested
replies are consolidated on a daily cycle into **AI Learnings** — plain-English rules you
can read, edit, enable, disable, or write yourself. Each rule is tagged **Outbound**,
**Reply Handling**, or both. Once you edit or author a rule, the AI stops modifying it.

### Where does this end and AI SDR begin?

The writing assistant produces words on demand: a sequence from a prompt, copy for a step,
a draft for a thread. [AI SDR](/capabilities/ai-sdr) is the standing motion — it derives
[offers](/specifications/offers) from your business, applies a
[playbook](/specifications/playbooks) to every generated message, sources and scores
contacts, personalizes per prospect, answers from a
[knowledge base](/specifications/knowledge-bases), and re-engages. If you are writing the
strategy and want help with the sentences, this page is the right one.

## Related

- [AI SDR](/capabilities/ai-sdr) — the full AI SDR motion; [The AI SDR model](/learn/ai-sdr-model) explains the model
- [Build an AI SDR](/workflows/build-an-ai-sdr) — the workflow where generated copy, playbooks, and approvals come together
- [Playbooks](/specifications/playbooks) — reusable message-generation instructions, the durable form of a prompt
- [Knowledge bases](/specifications/knowledge-bases) — what grounds AI replies in AI SDR sequences
- [Reply management](/capabilities/reply-management) · [Inbox](/specifications/inbox) · [Process replies](/how-to/process-replies)
- [Create a sequence](/how-to/create-a-sequence) · [Sequence steps](/specifications/sequence-steps) · [Templates and variables](/specifications/templates-and-variables)
- [Meetings](/capabilities/meetings) — the calendar side of meeting-intent handling
- [Sequence design](/learn/sequence-design) — what to ask the assistant for

## Build with Reply

- REST API: [docs.reply.io](https://docs.reply.io/api-reference/introduction) — check the reference for available AI and sequence operations
- MCP: [agents.reply.io/mcp](https://agents.reply.io/mcp) — agent access to sequences, inbox threads, and Jason AI configuration