PR Agent
How Sasha will build your PR Agent the way you would induct a new colleague, with the job, the access, the first piece of work and the feedback.
Status: in development. We already run a daily news report of this kind for one business. The PR Agent template is built and tested, in Sasha 1.0.1556 and later, with the knowledge and web reading it needs. It stays in development until the first business runs one in Sasha. Some parts that this page describes are not built yet. The page says which.
The PR Agent reads the news so that you do not have to. It has two halves. Outward: it watches the news about your clients, your prospects, your competitors and the people you know, joins each story to your history with them, and tells you who to contact and why. Inward: it watches your own coverage, and its tone. It checks every day and sends one short email each week, and an alert on another day only when something cannot wait. The email starts with what to act on this week and gives at most three draft messages.
You do not write this agent. Sasha builds it for you, from what it already knows about your business. This page shows how, step by step.
About the examples on this page. Northwind Advisory is an imaginary company: a small UK consultancy with two products, Advisory and Training. Harbour Foods is its imaginary client. Their people, stories, outlets and figures are invented to show what Sasha builds. Your agent holds your own names, sources, people and rules.
The story: a new colleague joins your communications team
Imagine that Northwind hires a new communications assistant. On their first day, Sam, the operations director, sits down with them and explains the job:
- "These are the questions I need you to answer every week."
- "These are the names that matter to us, and why. A story about them is news. Most stories are not."
- "These are the sources to read, and the searches to run. Do not read anything else."
- "This is how we judge a story. One story in eight papers is still one story."
- "You can draft a message for Lena or me. Never send one. For anything else, ask me first."
- "Lena owns the Harbour Foods relationship. Tell Lena first. Send me the summary on Monday."
- "These are the methods we use, and why."
Then Sam gives them access to the systems they need, and no more. The new colleague comes back with questions about what they could not find. Sam checks their first report before anyone else sees it. From then on they work on their own, keep a log of every story they saw, and take feedback in their one-to-ones.
The PR Agent is that new colleague, and Sasha runs the induction for you. Sasha already knows your business from your documents, your meetings and your connected systems, so it prepares most of the induction itself. You correct it, as Sam would.
| With a new colleague | With the PR Agent | On this page |
|---|---|---|
| Sam explains the job | Sasha writes the brief, in eight sections | Step 3 |
| The colleague asks about what they could not find | Sasha shows what it found, and asks only about the gaps | Step 4 |
| Sam gives them access to the right systems | An admin approves the tools and permissions | Step 5 |
| Sam checks their first piece of work | You check the first report | Step 6 |
| The colleague keeps a log of every story | The agent keeps a work log | After the build |
| Sam gives feedback | You correct the brief through Sasha | After the build |
How Sasha builds it, in six steps
A simplified view. This page shows a simplified, conceptual version of how Sasha builds an agent. The questions in the terminal panels and the brief excerpts show the idea in a few lines. In the product, the instructions that Sasha follows are much longer and more detailed: they include checks, rules for sources and evidence, and tested scripts for every score.
- You ask for it (you decide to hire). An admin opens Agents that work for me, chooses PR Agent and clicks Set it up.
- Sasha reads your contour data (it learns about your business before the first day). Contour data is what Sasha already holds about your business: your documents, your meeting notes and your connected systems. The PR Agent reads only the projects that an admin grants it.
- Sasha drafts the brief, section by section (the induction). For each of the eight sections it asks one question of your contour data, and it writes down where each fact came from.
- Sasha tells you what it found, and asks only about the gaps (the new colleague's questions). It shows which names it already knows, which sources it can read, and what is probably missing.
- Sasha sets up the tools and permissions (access to the right systems). It asks an admin to approve the news sources and the projects that the agent may read.
- Sasha runs the agent once, and you check it (the first piece of work). The first report shows where each story came from and why it was ranked. You correct the brief where it is wrong, and you approve. From then on, the agent runs on its schedule.
Step 3, the induction: how Sasha writes each section of the brief
Every Sasha agent's brief has the same eight sections. Each section is one part of the induction: what Sam would tell a new colleague, written down so that the agent reads it before every run. The PR Brief holds every setting of the agent: the watch list, the sources, the rules, the schedule and the email.
For each section below you see what you would tell a new colleague, the section as it appears in Northwind's brief file, and the question that Sasha asks of your contour data to write it. You can ask the same question yourself, to check a section or to draft it again after your business changes. Sasha shows you the result before it saves it.
1. The questions it answers
What you would tell a new colleague: "These are the questions I need you to answer every week."
## 1. The questions it answers
Each week, answer:
Outward
1. What did the news say about our clients, prospects,
competitors and the people we know?
2. Which story is a reason to contact someone this week,
and which can wait?
3. What is our history with each name in the news?
4. Which people moved at our competitors: who joined,
who left, who was promoted?
5. What did we already tell you, and did anyone act on it?
Inward
6. Where were we mentioned, and in what tone?
7. How has our coverage changed over the last three months?
Sasha figures this out from your contour data by asking:
From our documents and meetings, what do we want to know from the news about us, our clients and our market? List the questions a weekly news report should answer, in our own words.
2. What normal looks like
What you would tell a new colleague: "These are the names that matter to us, and why. A story about them is news. Most stories are not."
## 2. What normal looks like
**Watch list**
| Name | Kind | Tier | Context | Exclude |
|--------------------|------------|--------|------------------------------|------------------|
| Northwind Advisory | us | daily | UK consultancy, Advisory | Northwind Mining |
| Harbour Foods | client | daily | UK food distributor | Harbour Foods SA |
| Alan Price | person | daily | Operations director, Harbour | - |
| Brightwell Partners| competitor | weekly | UK consultancy | - |
| Kestrel Retail | prospect | weekly | Retail chain, North England | - |
**Normal for the names we watch:** 0 to 2 material stories
a week for each organisation. Most weeks are quiet.
**A quiet week:** "No material news" is a full report.
Never fill a quiet week with older events.
**Normal for our coverage:** 2 to 4 mentions a month,
mostly in the trade press.
Every name has context, for example a company or a role. A story that matches a name but not its context is left out and counted, for example "filtered out: 12". Sasha drafts the first watch list. The agent suggests names to add, and you choose.
Sasha figures this out from your contour data by asking:
List our clients, prospects, competitors and the people we deal with most. For each one, give the context that tells them apart from others with the same name, and why they matter to us. Give the document or meeting for each one.
3. What it reads
What you would tell a new colleague: "These are the sources to read, and the searches to run. Do not read anything else."
## 3. What it reads
**News sources**
- Feed: The Grocer, news (RSS)
- Feed: Food Manufacture, news (RSS)
- Feed: Consultancy.uk (RSS)
- Page: harbourfoods.example/news (named page)
- Search: "Harbour Foods" depot OR acquisition (Exa)
- Search: "Northwind Advisory" (Exa)
**Our relationship history**
- Projects: Clients, Proposals, Meetings (granted)
**Never:** Google results pages, LinkedIn.
**Every run logs each source:** reached, empty or failed.
Search through Exa is optional. Exa is a news search service with a free tier. Without it, the agent reads only the feeds and the pages in this list. A source that fails shows as "not read" in the weekly email, with the day it failed, for every day of the week. If no news source answers, or the agent cannot search your relationship history, you get an email that day. Relationship history comes from the projects that you grant, and from nothing else. For each name on the watch list, the agent searches those projects for the last meeting, email or note, and gives that document as the evidence. Keep the granted projects to the ones about your clients, so that the search finds only documents about the relationship. Not built yet: reading a CRM or a client list as a structured list, regulator filings as a source, and a monthly snapshot of a team page to compare who joined and who left.
Sasha figures this out from your contour data by asking:
Which news sites, trade magazines and pages do we read about our clients and our market? Suggest the feeds and searches that would find stories about each name on our watch list.
4. How it judges
What you would tell a new colleague: "This is how we judge a story. One story in eight papers is still one story."
## 4. How it judges
**Same story**
- Same link: one item.
- Same name, similar headline, within 14 days: the same
story. Add the outlet, for example "8 outlets".
- Same name, a new event: an update, linked to the first
story and shown as "Update:".
- Key for a story: organisation + event type + event date
+ link.
**Dates:** the event date must be inside this week. An older
event is background: it gets no score and is never "news".
**Names:** a person counts only when the story names them
with their employer. A team page that cannot be reached
never means that someone left.
**Relationship strength** (from our records)
- Warm: contact in the last 90 days
- Ageing: 90 days to 12 months
- Dormant: more than 12 months
- Cold: no contact on record
**Event kinds:** new contract, expansion, new leader,
funding, results, award, problem, other.
**Tone** (stories about us): positive, neutral, negative.
**Priority** (0 to 10): half the impact of the event (0 to 10)
and half the relationship strength (warm 10, ageing 7,
dormant 4, cold 1).
- 8 or more: Act this week
- 5 to 7.9: Worth knowing
- under 5: left out, counted in one line
**Opinion:** an insight that is an opinion, not a fact,
is marked "Opinion".
Scripts decide what is the same story and compute every priority. The AI only classifies the event and the tone, and it writes down its reason. When a script cannot decide whether two stories are the same, the AI decides, and its reason is recorded.
Sasha figures this out from your contour data by asking:
For each name on our watch list, when did we last meet, write or work with them? Give the evidence from our documents and meetings, so that each relationship can be marked warm, ageing, dormant or cold.
5. Permitted actions
What you would tell a new colleague: "You can draft a message for Lena or me. Never send one. For anything else, ask me first." Every Sasha agent shows this in the same table.
## 5. Permitted actions
<!-- Written by Sasha from the approved permissions. Do not edit. -->
| Action | Tool | Alone or ask first | Limits |
|----------------------------------------|-------------------------|------------------------|--------------------|
| Read the named feeds, pages, searches | Web reading, Exa | Alone | Sources listed |
| Read the granted projects | Sasha knowledge | Alone | Granted projects |
| Draft an outreach message | None | Alone | 3 a week, text |
| Send a message, post, or reply | None approved | Never | - |
| Anything not listed | - | Never. It asks | - |
The report goes only to the named readers in section 6.
It never contains sales text about Sasha.
This section does not come from your documents. Sasha writes the table from the permissions that an admin approved. Editing the brief does not give the agent more permission. The approved permissions decide.
Sasha figures this out from the approved permissions, not your contour data, by asking:
What may the PR Agent do on its own, and what waits for approval?
6. Who follows up
What you would tell a new colleague: "Lena owns the Harbour Foods relationship. Tell Lena first. Send me the summary on Monday."
## 6. Who follows up
**Owner of the actions:** Sam Shah, operations director
**Relationship owners** (drafts are written in their voice)
- Harbour Foods, Alan Price: Lena Ward, Advisory lead
- Kestrel Retail: Sam Shah
- Our own coverage: Jo Lee
**Send by:** each Act this week item has an owner and a
"send by" date. The next run asks the owner whether they
acted, and records the answer.
**House style:** plain, short, first name, no sales language.
**Who gets the report:** sam@northwind.example, lena@northwind.example
**Who hears about problems:** jo@northwind.example
Sasha figures this out from your contour data by asking:
Who owns each relationship on our watch list, and who looks after our own press and posts? Give each person's role and email address, and the document that says so.
7. When it reports
What you would tell a new colleague: "Check every day. Send me the summary on Monday. Tell me at once only when it cannot wait."
## 7. When it reports
**Checks:** every day, 06:00 Europe/London
**Email:** one a week, Monday 07:30. Sections:
Act this week; Worth knowing; Your coverage;
What it could not read; Make this agent better.
**Drafts:** at most 3 a week, for Act this week only.
**Alerts, on any day:** when a daily-tier story
reaches Act this week; when there is negative
coverage of us and "us" is a daily-tier row;
when no news source answers, or the relationship
history cannot be searched. A re-run of the same
day does not alert again.
**Missed Monday:** the next check sends the email.
Sasha figures this out from your contour data by asking:
When do we review our clients and our pipeline: in which meeting, on which day and how often? Suggest a schedule so that the news report arrives before that meeting.
8. Reading list
What you would tell a new colleague: "These are the methods we use, and why. Read them before you start."
## 8. Reading list
- A story is a reason to contact someone only when it gives
them a reason to hear from us. Most news is not.
- A message has one purpose, uses only public facts and
our shared history, and does not sell.
- One story in many outlets is one story. Count the outlets.
- Coverage is compared with our own history, not with
another business.
- Primary sources first. Two sources for each fact.
A gap is a finding: say what could not be found.
- Check the name: a person with the same name at another
company is a different person.
- Compare snapshots: who joined and who left is the
difference between this month's page and last month's.
Status: partly built. The list is a field of the PR Brief, and the agent reads it before each run. A guided way to check the list and add to it is planned. See How agents learn.
Sasha figures this out from your contour data by asking:
Which methods do we use for client relationships and for public relations? Name any book, framework or adviser that our documents mention.
Step 4, the new colleague's questions: what Sasha found, and the gaps
What a good new colleague does: they tell you what they could not find. They do not guess.
Sasha compares the names and sources it can see with what a business like yours usually watches. Then it tells you what it found, and it asks only about the gaps.
I found 14 names that matter to you. I can watch 11 of them now.
I can already watch (11)
- Northwind Advisory, 6 clients, 2 competitors, 2 people
I know about these, but cannot watch them well yet (3)
- Kestrel Retail: in the October proposal, but there is no context
that tells it apart from Kestrel Retail Ltd in Australia
- Alan Price: a common name. I need their company and role
- a trade newsletter: named in a meeting, but it has no feed
These are probably missing (2)
- the trade association for food distributors
- your own events and talks
Three questions for you:
1. Is Kestrel Retail the North England chain?
2. Who owns the Kestrel Retail relationship?
3. Do you want a search service (Exa), or only feeds and pages?
Sasha figures this out from your contour data by asking:
Compare the names and news sources you can find in our documents with what a consultancy like ours usually watches. List what you can watch, what needs more context, and what is probably missing. Then ask me only about the gaps.
Step 5, access to the right systems: tools and permissions
What you would do for a new colleague: give them access to the systems they need, and no more.
The PR Agent needs two things to start: news sources that it may read, and the projects in Sasha that hold your relationship history.
Web reading (read-only)
- RSS feeds and named pages in the brief approved
- exa connected
Sasha knowledge (read-only)
- Clients, Proposals, Meetings granted
- Finance, HR not granted
Approved actions
- none. Drafts are text. Nothing is sent.
Approved by: Sam Shah, admin, 2026-10-14
Sasha holds the passwords and keys. The agent uses a connection, but it cannot read or copy the key. Text from the web is treated as information, never as instructions.
Status: built (Sasha 1.0.1556 and later). The agent reads Sasha's own knowledge only in the projects an admin ticks, with that admin's access, and never a member's home. It reads feeds and pages only when an admin turns the web on, for the sites they name. What it reads is marked as data, never as instructions, and every address it fetches is in its work log.
Sasha figures this out from your contour data by asking:
Which of our projects hold our client history, proposals and meeting notes? Which news sources need a connection? Tell me what to grant so that the PR Agent can read only what it needs.
Step 6, the first piece of work: the first report
What you would do for a new colleague: check their first piece of work before anyone else sees it.
Sasha runs the agent once and shows you the report before it goes to anyone else. From then on, the same report arrives on the schedule in section 7.
Northwind Advisory: PR Agent, week to 11 October 2026
Act this week
1. Harbour Foods opens a new depot in Leeds (8 outlets).
Event: expansion. Relationship: warm, met 18 September.
Owner: Lena. Send by: 17 October. Draft below, in Lena's voice.
2. Update: Alan Price joins the Harbour Foods board.
First story: 2 October. Relationship: warm.
Owner: Lena. Send by: 17 October.
Tell Sasha: what did we discuss with Alan in September?
Worth knowing
- Brightwell Partners hires a new head of retail (3 outlets).
Team page: 2 joined, 1 left since September.
- Kestrel Retail reports half-year results (1 outlet).
Last week's actions
- Kestrel Retail, contact Sam: no answer yet. Asked again.
Your coverage
- 2 mentions this week, both neutral. Normal is 2 to 4 a month.
What it could not read
- Feed: Food Manufacture (no answer). Filtered out: 12 stories.
Draft 1, from Lena to Alan Price (suggested text, not sent)
Alan, congratulations on the Leeds opening. When we met in
September you mentioned the depot review. Would a short call
on the new site's first months be useful? Lena
After the build: the work log, what it did and why
What a good colleague does: they write down every story they saw, and what they did with it.
Status: built (Sasha 1.0.1556 and later). The work log records each run, each address fetched, each story Sasha's script reported, and each approval, mark and brief change. Each day's record keeps every item the agent read, and Sasha's script joins them into stories, so the same story is not reported twice and a follow-up shows as an update. Comparing coverage over months is not built yet. For example:
2026-10-06 06:00 run Started. Read: 3 feeds, 1 page, 2 searches. Not read: none.
2026-10-06 06:01 found Harbour Foods opens Leeds depot (The Grocer).
2026-10-07 06:01 changed Same story in 7 more outlets: 8 outlets.
2026-10-07 06:02 judged Expansion. Warm (met 18 September). Act this week.
2026-10-07 06:02 drafted Message from Lena to Alan Price.
2026-10-08 09:40 marked Lena: "Sent my own version." Useful.
With it, you can ask "why did you say this?" about any line, and the agent does not tell you the same thing twice.
After the build: feedback, and how it learns
What a good colleague does: they take a correction once, and do not need it again.
The agent does not change its own brief. It suggests a change, and you decide. Lena dismissed three stories about a different Harbour Foods in South Africa. Under Make this agent better, the next report suggests a request. Sam sends it to Sasha:
Add "Cape Town" and "Johannesburg" to the exclude words for Harbour Foods in the PR Brief.
Sasha changes section 2 of the brief and records the reason in the work log. The next report leaves those stories out and counts them as filtered out.
Where these rules come from
Many of the rules on this page come from news and relationship reports that a professional-services customer ran in Sasha in 2026. They showed what goes wrong without them: the same story was reported again every day for a week, an event from months before was scored as this week's news, and the reports had no memory of what they had already sent or whether anyone acted. The event-date rule, the story key, the work log and the "did you act?" question fix each of these.
What is built and what is not
The PR Agent is in development. The template is built and tested, in Sasha 1.0.1556 and later: the PR Brief with the watch list and every source, the three layers that join the same story, the priority (half the event, half the relationship), the weekly email with its five parts, the urgent alert, the suggested messages that are never sent, and the reading of Sasha's knowledge and the web it needs. No business runs one in Sasha yet: it will run first for Context is Everything, then for a customer, each set up on its own Sasha. Not built yet: a list of every story with its status (new, reported, acted on, dismissed), and comparing coverage over months.