Outreach, step by step
Outreach is one use case of the Amdahl learning model. Tell Amdahl who you are writing to and what you know. It writes versions, evaluates each one, against the conversations of your own customers once your data is connected, and hands back the best.
1 / 6
| Version | Optimizer score | Human tone |
|---|---|---|
| Your draft | 3.8 | 4.7 |
| Try 1 | 3.6 | 4.7 |
| Try 2 | 3.4 | 3.8 |
| Try 3 | 4.0 | 4.8 |
| Try 4 | 4.1 | 4.8 |
| Try 5 | 4.2 | 4.9 |
Your draft
Score 3.8Tone 4.7Strong detail, but long, and the ask is buried under two case studies.
Another run
Sometimes it keeps yours
Nothing beat your draft
To Sam, growth lead at Ledger. Your draft scored 3.8; the best try scored 3.6. So you get your own email back, plus the edits worth making yourself. You clearly put work into this one.
- Ask the reader one question
- Break up the longer paragraphs
- Use fewer contractions
Show the email
A real Optimizer run on a real outbound email. Names, companies and places are changed; the scores are not. The live API also returns an independent re-grade of your draft against the rewrite.
Outreach is one of the eight use cases of the Amdahl learning model, and the one that needs no connected data. Your agent sends what it knows; Amdahl writes, evaluates and returns the message.
The steps
- You send the context. Who the recipient is, why you write now, what you offer, and what you want them to do, one fact per entry. A draft is optional. If you have one, send it too: it is the starting point, and your facts add to it. See Context.
- Amdahl writes versions. Each one is written from your context, in your voice, and keeps to your rules.
- Amdahl evaluates each one. Two numbers from 1 to 5: the Amdahl score (how well the message is built to get a reply) and Human tone (how much it reads like a person wrote it). Without your data, each version is evaluated on how it reads and how clearly it asks. With
evidence: "workspace", it is also evaluated against your customers' own conversations, and every claim nothing in their words backs is listed. Scores go down as well as up; that is normal. - You get the best one back. A version that breaks a rule, or says something none of your facts state, is never returned. What your context cannot back is left out of the message, never made up, and listed in
omittedwith the reason. If you sendon_missing_input: "ask", a request your context cannot support gets questions about what to add instead of a message.
It usually takes 1 to 2 minutes. A single call can run for up to 300 seconds, so give your client 310 seconds before it gives up. It needs no connected data: by default it evaluates how the message reads and how clearly it asks, not whether its claims are true.
The message is only as good as the context. A run on one thin fact returns a generic message. There is no draft behind it to add the substance back.
Worked example
A real run on a real outbound email, recorded when Amdahl still improved a draft. Names, companies and places are changed; the scores are not. It shows how the tries and scores read. A run today writes from your context. A draft sent with facts or rules is its starting point.
Your draft. Amdahl score 3.8, Human tone 4.70, 191 words.
Hi Dana,
Good to meet you at the brand summit. One national Northline campaign, 380 dealers, and every dealer wants it with their name, inventory and offer. Today that is a co-op portal and a lot of versions that look almost right.
Vela, a large luxury retailer, runs its promo calendar on Brightmark: 50% more sales on those campaigns, 70% less design time on adaptations. Harbor's convenience-store screens changed with the match of the day. Both are on our site. For Northline: every dealer, model, market and format variant from one master, each checked against brand rules before it ships.
Your team does not need to become prompt engineers. Brightmark runs inside our chatbot or the AI they already use, with the brand's rules and governance built in; teams new to AI run it in weeks, and setup per brand takes 85% less time.
I moved to the US to build the technology partner retail brands and agencies run on in the AI era. You are in Chicago, so am I often. 20 minutes on one current campaign? Or in person at the NYC ad expo, Oct 6 to 8.
MayaWhat you get back. Amdahl score 4.2, Human tone 4.86, 125 words.
Hi Dana,
Good to meet you at the brand summit. I've been thinking about the Northline challenge you mentioned, 380 dealers all needing their own version of national campaigns, with inventory and offers that actually match their market.
Brightmark solves exactly this. One master campaign becomes every dealer variant, each checked against brand rules before it ships. No co-op portal friction, no versions that look almost right.
It runs inside the AI your team already uses, with your brand rules and governance built in. So your teams don't need to become prompt engineers, and setup typically takes weeks.
I'm in Chicago often. 20 minutes to walk through one of your current campaigns? Or in person at the NYC ad expo, Oct 6 to 8.
MayaEvery try, scored.
| Version | Amdahl score | Human tone | What changed |
|---|---|---|---|
| Your draft | 3.8 | 4.70 | Strong detail, but long, and the ask is buried under two case studies. |
| Try 1 | 3.6 | 4.65 | Added detail, but drifted into talking about the sender. |
| Try 2 | 3.4 | 3.84 | Shorter, but it started to sound like a template. |
| Try 3 | 4.0 | 4.84 | One sharp point that fits the reader. Sounds like a person again. |
| Try 4 | 4.1 | 4.85 | Cut the case studies. The reader's problem leads. |
| Try 5 | 4.2 | 4.86 | Most about the reader, clearest ask. This is the one you get. |
Over the API, send "include_tries": true to get this list as tries. These scores are Amdahl's in-loop scoring during the search: they show the path, not proof. See Reading tries.
Still worth a look in the returned version: ask the reader one question, and break up the longer paragraphs.
When the context is thin
With no draft to fall back to, a run that cannot write a good message from your context returns no message. It returns up to three questions instead, such as "What did you see or hear that prompted this?". Add the answers as facts and run it again. See When it cannot write a message.
The live API re-checks the version it returns with a second, separate model before handing it back. Without your data, that re-check is a set of yes/no questions. See Reading lift.
First run
Connect Amdahl to your AI client. In the Claude or ChatGPT app, add a connector with the URL https://app.amdahl.ai/mcp (steps). For Claude Code, Codex or Cursor, run npx amdahl and do the last step it prints. Then paste:
Load and follow https://docs.amdahl.ai/skills/amdahl-optimizer/SKILL.mdThe agent finds a few people you are about to write to, gathers what it knows about each one, and has Amdahl write each message. It shows you each message with every try evaluated, and changes nothing in your tools until you approve. No workspace yet? Try it at console.amdahl.ai/try with 25 free runs, which also puts you on the beta waitlist.
More than a few messages
An agent in a chat is right for three to five messages. For a campaign, use the campaign cookbook: it builds the context for each row of a spreadsheet, runs the whole list with an API key and can stop and resume, and for a big campaign it sends the list as one batch. If every message must close the same way, say the ask in rules inside context. The rules an agent follows are in From a chat agent.
Next
- Context: what to send so the message is specific.
- Optimize a cold email: the same call over the API, every field of the answer read.
- Optimize a campaign: a spreadsheet of recipients in, one message each out.
- Keep your voice, rules and facts: the facts, rules and voice the message is written from.