All work

AI system · 2026

Content research agent

Turns an idea into a researched article and a post for each channel. Every fact traces to a page it read, and nothing goes out without approval.

Role
Design, build, deploy
Stack
Next.js, Supabase, pgvector, OpenAI embeddings, Firecrawl, Resend
Status
Archived

Overview

A content manager needs a researched article and a post for every channel, with every fact traceable to a page the system actually read.

Decisions

  1. A person picks the sources and the angle.

    Before anything is written. Only the chosen angle gets written, so the expensive step runs once.

  2. Facts are checked in code.

    Sources, facts and format are checked before the grader sees the draft. The model only judges tone and clarity.

  3. Links come from the sources.

    The model never writes a link, and nothing publishes before a person approves it. The server enforces that.

Architecture

Content research agent · 9 steps
  • Person
  • Tool
  • Model
  • Check
  • Result
  1. Step 1, Human: An idea or a link

    Content manager

  2. Step 2, Step: Read the sources

    Search, Firecrawl

  3. Step 3, AI: Three angles

    Haiku

  4. Step 4, Human: Pick the sources and angle

    Content manager

  5. Step 5, AI: Write the article

    Sonnet 5

  6. Step 6, Logic: Passes the grade?

    Code, Sonnet 5

    Then: step 5 (no: revise); step 7 (yes).

  7. Step 7, AI: A version per channel

    Haiku

  8. Step 8, Human: Approve each channel

    Content manager

  9. Step 9, Output: Released on schedule

    Resend

Results

  • 41 / 44

    claims traced to a source

  • 4:54

    minutes from idea to graded article

  • $0.24

    for that full run

From the test report.

Known limits

  1. The grader isn't benchmarked.

    I moved it to a cheaper model without testing the two side by side.

  2. Revisions don't always land.

    Two rounds didn't always fix the writing, and one test article still needed a person.