The draft cites sources that don't say what Claude claims

Always read step 2's source list before approving the outline. Add: 'quote the exact sentence from each source you're relying on.' If it can't quote it, the claim isn't really sourced.

Every draft sounds the same and a bit robotic

You haven't given it a voice. Paste two paragraphs you wrote and liked, and say 'match this rhythm and word choice'. With Memory MCP, save it once as 'my writing voice' and it carries forward.

Exa returns thin or off-topic results

Your query is too broad. Make the angle narrower before researching — 'AI code review' is noise; 'why AI PR review misses logic bugs in unchanged files' returns real material.

Filesystem MCP refuses to write the file

It can only write inside the directories it was granted at install. Point your drafts folder at one of those paths, or re-install Filesystem MCP scoped to your drafts directory.

How long does it take to daily blog draft pipeline: research to draft with claude?

About 25 minutes (10 to set up the first time, 15 per draft after). Difficulty: beginner.

What do I need for this?

Exa. Turn them on in Claude or ChatGPT, then run the prompts — no code.

Do I need to know how to code?

No. Turn the connectors on in your AI's settings and paste the prompts in plain words.

Recipe

Daily blog draft pipeline: research to draft with Claude

Go from a one-line topic to a researched, sourced first draft saved to disk — before your coffee is cold, every day.

25 minutes (10 to set up the first time, 15 per draft after) beginner 3 connectors for Creators who publish weekly to Instagram, LinkedIn, Twitter, Threads, Telegram

What you'll need

Exa Searches the live web for current, citable sources so the draft is grounded, not hallucinated.

Required

Filesystem Writes the finished draft to your drafts folder as a dated markdown file.

Optional

Memory Remembers your writing voice and structure so every draft sounds like you, not like a model.

Optional

Do this, in order

1
State the topic and the reader

Topic for today's draft: [your one-line topic]. Reader: [who they are and what they already know]. Goal of the post: [teach / persuade / announce]. Target length: 800-1100 words. Don't write yet — first tell me the angle you'd take and why.

You'll get: Claude proposes one specific angle and a reason. You approve it or push back ('too generic, make it about the failure case').
2
Research with real sources

Use Exa to find 4-6 current, credible sources on this angle from the last 18 months. For each: the claim it supports, the URL, and one line on why it's trustworthy. Skip SEO listicles and anything undated.

You'll get: A short source list with URLs and a one-line credibility note each. This is the spine of the draft — check it before writing.
3
Outline, then draft

Outline the post: hook, 3-4 sections, takeaway. Show the outline. Once I approve, write the full draft in my voice — short paragraphs, concrete examples, no 'in today's fast-paced world' openers. Cite the sources inline as [1], [2].

You'll get: An outline first (you approve), then a complete draft with inline citation markers and a sources list at the bottom.
4
Save it with a dated filename

Save the draft to my drafts folder as YYYY-MM-DD-[slug].md with frontmatter (title, date, status: draft). Print the full path so I can open it.

You'll get: A markdown file written to disk via Filesystem MCP, with a path you can open in your editor. Without Filesystem MCP, Claude prints the full draft for you to paste.

You're done when

A 800-1100 word first draft, grounded in 4-6 real sources you can verify, written in your voice and saved to disk as a dated file — produced in 15 minutes from a one-line topic. The blank page never happens. You edit a draft instead of starting one.

Why this workflow exists

Consistent publishing dies on two things: the blank page and the research tab-storm. You know the topic. You just don’t have a draft, and getting one means an hour of opening sources, skimming, and forcing the first paragraph. Most days that hour doesn’t exist, so the post doesn’t either.

This workflow removes both. Exa MCP gives Claude live, citable sources instead of a model guessing. Filesystem MCP lands the result in your drafts folder as a real file. Your job moves from write a post to edit this draft and check its sources — a much smaller, much more reliable task that fits in a real morning.

The first run takes ~25 minutes because you’re installing servers and teaching Claude your voice. After that it’s ~15 minutes from a one-line topic to a saved draft.

Install in 10 minutes

You need Exa MCP for research. Filesystem MCP and Memory MCP are optional but they’re what make this a daily habit instead of a one-off.

1. Install Exa MCP

claude mcp add exa -e EXA_API_KEY=<your-key> -- npx -y exa-mcp-server

Get an API key from the Exa dashboard. The maintained command and free-tier limits are on the Exa MCP page. Verify with /mcp — exa should be connected.

2. (Optional) Install Filesystem MCP

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem <path-to-your-drafts-folder>

Replace the last argument with the folder you want drafts written to. Filesystem MCP can only touch directories you grant it here — see the Filesystem MCP page.

3. (Optional) Install Memory MCP

claude mcp add memory -- npx -y @modelcontextprotocol/server-memory

No env vars. This holds your writing voice so drafts stop sounding generic after the first calibration.

The first run

Run prompt 1 and actually argue with the angle — the angle decides whether the post is worth publishing, so don’t rubber-stamp it. Run prompt 2 and read the sources: this is the one step you can’t skip, because a grounded draft is only as good as the spine under it.

Approve the outline in prompt 3, then let it draft. The first draft will be 70% there — that’s the point. You’re editing, not writing. Prompt 4 saves it; from then on you open your editor to a file, never a blank page.

What success looks like

You typed one line at 9am. By 9:15 there’s a dated markdown file in your drafts folder: 900 words, your voice, six real sources you can click and verify. You spend the next 20 minutes editing a draft instead of the next 90 minutes building one. Tomorrow you do it again.

Going further

Series mode. “This is part 2 of a 4-part series — keep the voice and structure consistent with the file I saved yesterday.”
Repurpose downstream. Once the draft is saved, pair this with the Instagram carousel recipe to turn the same post into a carousel.
Source-quality bar. Tell Claude to refuse to cite anything without a clear author and date — it raises the floor of every draft automatically.

Exa MCP — live, citable web research
Filesystem MCP — writes the draft to disk
Memory MCP — holds your writing voice across drafts

If something breaks

The draft cites sources that don't say what Claude claims: Fix: Always read step 2's source list before approving the outline. Add: 'quote the exact sentence from each source you're relying on.' If it can't quote it, the claim isn't really sourced.
Every draft sounds the same and a bit robotic: Fix: You haven't given it a voice. Paste two paragraphs you wrote and liked, and say 'match this rhythm and word choice'. With Memory MCP, save it once as 'my writing voice' and it carries forward.
Exa returns thin or off-topic results: Fix: Your query is too broad. Make the angle narrower before researching — 'AI code review' is noise; 'why AI PR review misses logic bugs in unchanged files' returns real material.
Filesystem MCP refuses to write the file: Fix: It can only write inside the directories it was granted at install. Point your drafts folder at one of those paths, or re-install Filesystem MCP scoped to your drafts directory.

Exa Filesystem Memory