LinkedIn outreach selection

Summary: A weekly workflow where the user dumps their LinkedIn connections export into a project, the AI searches across those connections using accumulated business context, and surfaces a batch of contacts worth reaching out to that week — with a suggested angle for each. The user can optionally add a focus for the week (e.g., "people in HR tech," "founders raising right now") to narrow the selection. The user reviews each suggestion, edits the draft message, and sends.

What makes this automation-level: The AI handles search, ranking, and first-draft messaging across hundreds of contacts without the user driving each evaluation. The user reviews the batch and edits drafts before sending — but the work of identifying who fits and why is delegated. Without the AI, this would require manually scrolling through connections and remembering context for each one, which doesn't scale and rarely happens consistently.

Map

The recurring workflow is weekly outbound — picking who to reach out to and writing the opener. It decomposes into: export LinkedIn connections, decide whether this week has a specific focus, search through the connection list using business context, score connections by relevance, draft an opening message tailored to each contact, send, log responses.

The search, scoring, and drafting steps were all classified as good AI candidates: they're pattern-matching tasks against a known reference (the business context), the volume is too high to do thoroughly by hand, and the output is easy to verify because the user reads every draft before it goes out.

Test

Initial test was a single prompt asking the AI to find ten relevant connections from a CSV given a one-paragraph description of the business. The output was workable but generic — the suggested angles were variations on "we should chat about AI."

The breakthrough came from feeding the AI richer business context: not just what the business does, but recent work examples, who the ideal customer profile actually looks like, what kinds of conversations have led to real engagements, and the user's own voice in past outreach. With that context loaded, the suggested contacts and angles got materially sharper — specific reasons each person was a fit, and angle suggestions that connected to actual recent work.

Integrate

Built as a project-based assistant with three loaded artefacts: the LinkedIn connections export (kept current by re-exporting periodically), a business context file describing what the user does, ICP, and recent work, and a folder of past outreach messages organised by recipient and outcome.

The weekly trigger is a single instruction: "find twenty contacts to reach out to this week" — optionally with a focus appended. The assistant searches across the connections list, ranks by fit, drafts a tailored opener for each, and presents the batch. The user works through the batch, editing or rejecting each draft before sending. Final-version drafts get saved back into the past-messages folder as the new database entry.

Compound

Two compound mechanisms run in parallel and reinforce each other.

The first is business context refinement. As the vault accumulates more material — case studies, completed engagements, refined ICP descriptions, lessons from won and lost conversations — the assistant has more to draw from when scoring fit and suggesting angles. The improvement is invisible week-to-week but compounds: contact suggestions a year in are noticeably more targeted than they were at the start, even with the same prompt, because the underlying context is richer.

The second is the past-messages database. Every final-draft message that gets sent is saved back into the project. Over time this accumulates into a reference library the assistant uses when drafting new openers — matched to relationship type, industry, and angle. New drafts inherit the user's voice more accurately because there are concrete examples to anchor on, and successful angles that recur in the database start showing up in new suggestions. Corrections the user makes in the editing pass also feed this loop: a draft the user rewrites significantly becomes a higher-quality reference than the AI's original.

The two mechanisms compound together because better business context improves which contacts get surfaced, and better past-message references improve how those contacts get approached.

agent-use-case-evaluation — This workflow passes the filter: high recurring volume, verifiable output (every draft is reviewed), bounded inputs (a known connection list and a known business context), and the cost of a bad single run is small (the user catches it before sending)
compound-engineering — The two-mechanism compound pattern (context refinement plus reference accumulation) is a strong instance of how a workflow can get materially better over time without rebuilding the underlying prompt
ai-output-verification — Outreach drafts are easy to verify because the user reads every one before sending; tone and accuracy issues are caught in the editing pass and fed back as future references