TemidireOpen for projects
All work

Case 02 / 02Koya Talent cohort, week 4, 2026

Content research agent

Turns a raw idea or a URL into a cited article plus LinkedIn, X and newsletter posts, with a person signing off before anything goes out.

Open the live app View the code
Client
Koya Talent cohort, week 4
Year
2026
Role
Solo build, design to deploy
Timeline
1 week
Result
One idea in, four channel-ready pieces out, human-approved
Runs per content request
Submit an idea or a URLResearch and rank sourcesDraft with citationsEvaluate and reviseAdapt, approve, publish
8rubric criteria every draft is graded on
14citations across 6 sources in the sample article
3channels adapted from one approved draft

01The problem

The content team brainstormed, researched, wrote SEO articles, rewrote them for each channel, reviewed and published, all by hand. It worked, but every extra piece cost the same manual time, and tone and accuracy got harder to keep consistent as volume grew.

02The approach

A Next.js and Supabase app that runs the middle of the pipeline. Firecrawl researches the idea or reads the given URL, the sources are ranked, Claude writes a cited draft, grades it against an eight-point rubric and revises the weak sections, then adapts it for each channel. A person approves before it reaches the publishing queue.

Claude APIFFirecrawlSupabaseNext.jsResendVercel

03How it works

Step 01

Submit an idea or a URL

A raw idea or a source URL, plus the target audience. Gibberish, unreachable links and paywalled or login-only pages are caught before anything is written.

Step 02

Research and rank sources

Firecrawl searches the idea or reads the URL. Sources are ranked for relevance, and the member picks which ones the draft is allowed to use.

Step 03

Draft with citations

Claude writes an SEO article where every claim is tied to one of the chosen sources, so the reviewer can see exactly what informed it.

Step 04

Evaluate and revise

The draft is graded on tone, clarity, SEO fit, audience fit, completeness, topic relevance, source grounding and factual consistency. Weak sections are revised, up to three passes, with every version kept.

Step 05

Adapt, approve, publish

LinkedIn, X and newsletter versions are checked against each platform's formatting rules. Nothing moves until a person approves it, then it goes to a publishing queue to publish now or schedule.

04Before and after

Before

  • Ideas researched by hand
  • Articles drafted from scratch
  • Each channel rewritten separately
  • Quality and tone checked by eye

After

  • Sources found, ranked and cited
  • Drafts graded on 8 criteria, with version history
  • Channel posts checked against formatting rules
  • Nothing goes out until a person approves it

05See it run

Channel adaptation with an X post preview that fits the formatting rules
Channel posts, checked
A new content request with its idea and target audience
Where a request starts

06Proof it holds up

  • Idea-only and URL-based requests both produce grounded drafts.
  • A members-only URL that returned a login stub is now caught before a draft is written from it.
  • Formatting checks flag posts that break platform rules, such as too many emoji on LinkedIn.
  • Approval locks the request; comments go both ways between reviewer and writer.
One idea in, four channel-ready pieces out, human-approved

07What I learned

I started by generating two article options per request. Carrying both through evaluation, revision and channel adaptation roughly doubled the cost of every later stage, for a comparison nobody used. Dropping to one draft with section-level revision made it cheaper and better.

Live app Source code Walkthrough on Loom
Next caseLead research agent