Most of what your agent says is shaped by four settings: your offering, your customers' pain points, a tone, and a length.
Additional outreach instructions is where everything else goes. It is a free text field, and it is the difference between an agent that sounds like generic outbound and one that sounds like it works at your company.
Where to find it
Open any agent, go to Settings, then Outreach settings. The field sits below the outreach threshold and above Goals, labelled Additional Outreach Instructions, with the note "Any additional rules the agent should follow when writing."
Instructions apply to that agent only. This matters more than it sounds, and it is the reason the answer to "how do I handle three different regions" is usually three agents rather than one very long instruction block.
What this field controls
Anything the agent writes. That means outbound messages and follow-ups on every mode, and replies as well if the agent is on Full Autopilot.
There is one scoping rule worth knowing. Inside each message step of your workflow, you choose between "Use your own message" and "AI Contextual Outreach". Instructions shape what the agent writes, so they apply in full to steps set to AI Contextual Outreach. On a step where you wrote the copy yourself, your copy is sent as you wrote it. If your instructions do not seem to be taking effect, check which mode your message steps are on.
What belongs somewhere else
The most common mistake with this field is using it for things the product already handles properly elsewhere. An instruction is a suggestion to a language model. A setting is a rule. Prefer the setting.
If you want to | Do not write an instruction | Use this instead |
Never contact certain people or companies | "Never message anyone at Acme" | Blacklist, under your account name at the bottom left. Six tabs, and a contains condition catches every variation of a company name in one line. |
Only contact people who fit a profile | "Only message VPs at companies over 200 people" | Targeting, under Settings. Job titles, locations, industries, company types, and company sizes are real filters. |
Only contact strong matches | "Only reach out if they seem like a good fit" | The outreach threshold, under Settings then Outreach settings. It is a hard gate, not a judgement call. |
Change how long messages are | "Keep it short" | Message length. Setting it to concise and then also writing "keep it short" is redundant at best and contradictory at worst. |
Change the overall register | "Be professional" | Tone of voice. Professional, friendly, or bold. |
Describe what you sell | Your whole pitch, pasted in | Agent profile, under Settings. Your Offering and Customer Pain Points are what every message is built from. |
Share a booking link | "Send them my Calendly" | The link field next to Goals. The agent shares it when the prospect is ready. |
Stop contacting someone another agent is already talking to | "Do not double up" | Leave Include Locked Profiles unchecked on the prospect source. It excludes anyone another agent messaged in the last 14 days. |
Use this field for the things none of those settings can express: judgement, nuance, restrictions, and the specific way your company talks.
How to write an instruction that works
One rule per line. A paragraph containing five rules gets followed less reliably than five separate lines.
Be specific rather than general. "Avoid corporate jargon" is weaker than "never use the words leverage, synergy, or circle back".
Say what to do, not only what to avoid. A rule with no alternative leaves the agent to invent one.
Make it checkable. If you cannot tell by reading a sent message whether the rule was followed, the agent cannot either.
Do not fight your other settings. If tone is set to bold and your instructions say be understated, you will get something inconsistent.
Start with five or six rules. Add more only when you have read real messages and found a gap.
What to put in it
Brand voice and style
“Sound friendly, concise, and confident. Avoid corporate jargon.”
“Keep replies warm but professional. Use short paragraphs.”
“Use a consultative tone, never pushy or salesy.”
“Never open with "I hope this message finds you well" or any variation of it.”
“Write at a plain reading level. No semicolons, no em dashes.”
“Never use emoji.”
Referencing the signal that surfaced the prospect
This is the one most people miss, and it is the one that makes signal-led outreach feel different from a mail merge. Your agent knows why a prospect appeared: a hiring spike, a funding round, a technology change, a review surge, a podcast appearance.
How openly it says so is your call, and it should be a deliberate one.
“Reference the signal that surfaced this prospect in the first message, but describe it in plain language rather than naming a data source.”
“Never say "I noticed you are hiring" directly. Instead, refer to what growth at that stage usually creates.”
“If the signal is a competitor review complaint, never name the competitor. Refer to the problem, not the vendor.”
“If the signal is more than 30 days old, do not treat it as news.”
Region, language, and local custom
“For prospects in Germany, maintain a formal tone and avoid first names unless they use mine first.”
“For Japanese prospects, keep messages highly polite and avoid overly casual statements.”
“For UK prospects, avoid American idioms and keep the tone understated rather than enthusiastic.”
“Reply in the language the prospect wrote in. If they wrote in French, reply in French.”
“Never suggest a specific meeting time. Ask what works in their time zone.”
If a single agent is covering several regions, consider splitting it. Regional rules stack up quickly, and an agent with one clear set of conventions outperforms one carrying four.
Objection handling and relationship selling
“If the prospect seems hesitant, offer something small rather than pushing. For example: happy to share a quick example if that would help.”
“If they say it is not a priority right now, ask when would be a better time to revisit and leave it there.”
“If they say they already use a competitor, do not compare. Ask what is working well and what is not.”
“Never handle the same objection twice. If they decline a second time, thank them and stop.”
Reply handling rules
These matter most on Full Autopilot, where the agent answers without you. They are still worth setting on the other modes, because they shape follow-ups.
“If the reply is an out-of-office message, do not treat it as a response. Wait and continue the sequence.”
“If they say they are not the right person, thank them and ask who is.”
“If they ask a question about security, legal terms, or pricing detail, do not answer. Say I will follow up personally and stop replying.”
“Never invent a case study, customer name, or statistic. If asked for proof and there is nothing in the agent profile, say I will send it over separately.”
Compliance and claims
“Never promise results or guarantees.”
“Avoid discussing pricing unless the prospect asks.”
“Never claim we are used by a named company.”
“If a prospect asks to be removed, confirm in one line and do not send anything further.”
Add anything your industry requires. Regulated sectors usually have language that must appear, or must not.
Industry and persona
“For technical leads, reference integrations and APIs when relevant.”
“For HR prospects, emphasise time savings and candidate quality.”
“For finance prospects, lead with risk and control rather than speed.”
Meeting and next-step logistics
“Only share the booking link once the prospect has expressed interest. Never in the first message.”
“If they suggest a call, agree and share the link rather than proposing a time.”
“If they ask for a demo, confirm and share the link in the same message.”
Cadence and stopping
“Never send more than two follow-ups without a reply.”
“If a prospect has not replied after the second follow-up, do not reference the earlier messages again.”
“Never send on the same day as the previous message.”
Creative uses
The field also takes instructions that are harder to categorise, and these are often the ones that change how the agent reads.
Persona: "Sound like a senior strategist, not a junior rep."
Humour: "Add light, subtle humour only when the prospect has used it first."
Confidence: "Keep the tone assertive but never aggressive."
Scarcity: "If they ask about availability, mention that onboarding slots are limited, without pressure."
Qualification: "If the prospect is outside the US or EU, ask what region they sell to before proposing next steps."
Self-awareness: "If asked whether this is automated, say yes plainly and offer to continue the conversation personally."
That last one is worth thinking about before someone asks.
How autopilot behaves in your inbox
Once an agent is on Full Autopilot, it answers on your behalf inside your LinkedIn inbox.
Messages the agent sent are marked so you can tell them from your own.
A banner sits above the reply box reading "The agent is handling this conversation", with a Take over button. Click it and the agent stops replying on that thread and hands it to you.
The gear icon next to the banner opens the settings of the agent that owns the conversation, which is the fastest way to fix an instruction that is not landing.
The agent replies within one to four hours of the prospect's message rather than instantly, so the timing reads as human.
You can take over at any point, including part-way through a conversation the agent started. On Outbound Autopilot and Co-pilot the agent never replies at all, so every response is yours from the first one.
Testing and refining
Instructions are guidance, not code, and the only way to know whether they landed is to read what went out.
Read the actual messages. Open the Inbox and read what the agent sent, not what you think it will send.
Change one rule at a time. Changing four at once tells you nothing about which one worked.
Use variants for the big questions. Inside a message step you can add a second version and Alsona splits sends between them. This is a cleaner test than rewriting instructions and hoping.
Watch the reply sentiment split on the dashboard. If negative replies rise after an instruction change, that is your answer.
Check the breakdown by signal source on the Intent Signals tab. If one type of signal replies well and another does not, the fix is often an instruction about how to reference that signal.
Final tips
Be specific rather than generic. Every vague instruction is a decision you have handed back to the model.
Split agents rather than stacking rules. Separate agents per region, per persona, or per offer will beat one agent carrying every exception.
Include objection handling. It is the part most people leave out and the part that decides whether a reply becomes a conversation.
Keep the core message simple. Your Offering and Customer Pain Points are doing most of the work. Instructions refine, they do not replace.
Revisit after the first fifty replies. You will find things you could not have predicted.
Test tone as well as instructions. Professional, friendly, and bold produce genuinely different messages from the same instruction set.
Write down why you added each rule. Six months on, an instruction block nobody can explain is one nobody will dare change.