# Build an Evidence-Led LinkedIn Agent

A documentation-first system for researching, drafting, and reviewing LinkedIn content while a human retains control of every public action.

Published: 2026-08-19
Canonical: https://darrylwong.me/posts/build-evidence-led-linkedin-agent
Topics: LinkedIn, Content Operations, AI Agents, Human-in-the-Loop

## Prepare stronger work without automating engagement

A useful LinkedIn agent should help with strategy, research, post drafting, comment preparation, and scheduled reviews. It should not publish, comment, react, message, scrape, or bypass platform controls on its own.

The operating boundary is simple: the system prepares review-ready work, while a human applies editorial judgement and separately approves each public action at the time it occurs.

## Build the system in seven inspectable layers

The guide separates project architecture, source knowledge, positioning, post generation, comment preparation, scheduling, and human approval. Each layer produces a durable artifact that another operator can inspect and explain.

Captured evidence stays separate from maintained synthesis. Positioning is written as an audience, problem, earned proof, and boundary statement before a drafting skill is allowed to turn research into content.

## Make every draft carry its own review context

A post package should include its objective, intended audience, hook, copy, claims requiring verification, image direction, alt text, hashtags, call to action, and publication checklist. This makes review more precise than approving copy in isolation.

A comment package should include the source, relationship context, fit score, contribution type, claims to verify, and a recommendation. The agent should return no draft when it has nothing credible or useful to add.

## Measure fit and substance instead of activity

Comment opportunities can be scored on audience fit, subject fit, relationship value, credible contribution, and timeliness. Generic praise, summary-only replies, self-promotion, and unverifiable claims should fail the selection step.

Useful review looks beyond impressions. Audience fit, saves and sends, substantive comments, profile actions, and qualified outcomes provide better feedback for the next drafting cycle.

## Schedule preparation, not authority

Recurring tasks can gather permitted context and prepare drafts after the skills have passed manual tests. A schedule does not grant standing permission to engage publicly, and an earlier approval should not be reused for a later action.

The living Gist contains the seven-layer model, eight-step build sequence, required outputs, safety rules, an exact configurable implementation prompt, completion check, and links to current platform guidance that should be rechecked as policies change.

## Sources

- [Read the one-page LinkedIn agent guide](https://gist.github.com/oruenboi/c83f6fc52f82e83081dcab80bab0e4c7)
- [Build a Markdown second brain in Codex](https://gist.github.com/oruenboi/4493518384a206a62afb3ce6fe3a0e26)
- [Adopt Vercel Eve architecture in Codex](https://gist.github.com/oruenboi/0b62aa40908d9562e4385deb850c2852)