Step 5: Approve, Rewrite, or Kill
Look at each draft. Three choices.
Approve. It sounds like you. Send it.
Rewrite. It is close. Fix the parts that do not sound like you, add the specific detail Claude missed. Send it.
Kill. The thread is dead or the draft is too generic to save. Archive the thread or mark it as PARKED.
Track this. After a week you will see a pattern.
If Claude's first drafts are getting killed more than 30% of the time, your voice file is too thin. Go back and add more examples of how you really talk.
If approval rate is above 70%, your voice file is doing its job.
Step 6: Set the Follow-Up Reminders
Most DM stalls happen here. You replied. They replied. You forgot to reply back. Two weeks later the thread is cold.
Type the prompt:
For every thread where I just sent a reply, create a Kondo reminder:
- If I asked a question: remind me in 3 days if they haven't replied
- If I suggested a call: remind me in 5 days if they haven't booked
- If I shared something useful: remind me in 7 days to follow up
Add a note to each reminder so future-me knows what we were discussing.
Claude writes the reminders into Kondo. Kondo nudges you when the time comes. You never forget a thread again.
Step 7: End-of-Week Review
Friday afternoon. Pull a week summary.
Type the prompt:
Summarize my LinkedIn DM activity from the last 7 days:
- Total threads I replied to
- New threads that became qualified conversations (question asked, call suggested)
- Threads I killed and why
- Any prospect who asked for a call
- My approximate reply rate on cold first messages I sent
Output it as a one-page summary I can paste into my weekly review.
Claude writes the summary. Ten minutes of reading tells you what worked and what did not. That one screen replaces an hour of scrolling back through threads.
What I'd Do Differently
Build the voice file first, not last. The first week I ran this stack my voice file was four lines. Claude drafts sounded like a LinkedIn content coach. I rewrote every one.
I finally sat down and wrote 3 pages of how I really talk. 20 real example sentences. The next week's drafts were 70% sendable.
Start with Kondo before writing any Claude prompts. I spent two days trying to get Claude to scrape LinkedIn directly. Then I found Kondo.
Kondo does the integration part. Claude does the thinking part. Do not try to make Claude do both.
Keep the approval loop forever. My reply rate on AI-drafted DMs without human review would drop to zero in a month. Prospects know.
The human in the loop is not slowing you down. It is the only reason the messages work.
The Stack
| Tool | What It Does | Cost |
|---|---|---|
| Kondo | Pulls and tags the LinkedIn inbox, stores thread history, sets reminders | Free tier to start |
| Claude Code | Reads the inbox, drafts replies, sorts priority, writes summaries | $20/mo standard, $200/mo max |
| The platform everything runs on | Free | |
| Voice file | One markdown file of your tone and examples | Zero, you write it |
Total: $20 a month plus whatever Kondo tier you pick. I run on $20 Claude Code and Kondo free for the first three months.
Files
~/dm-workflow/
voice.md (your tone, banned words, example sentences)
prompts/
pull-inbox.md (Step 1 prompt)
triage.md (Step 2 prompt)
priority-5.md (Step 3 prompt)
draft-replies.md (Step 4 prompt)
set-reminders.md (Step 6 prompt)
weekly-review.md (Step 7 prompt)
logs/
weekly-YYYY-MM-DD.md (Friday summaries, one per week)
Seven prompts. One voice file. One folder. That's it.
Frequently Asked Questions
How long does setup take?
One afternoon if you already have Kondo and Claude Code installed. Two or three days if you are starting from zero.
The longest step is writing the voice file. The rest is connecting accounts and saving the seven prompts.
Can Claude send my DMs automatically?
No, and it should not. Automated sending is what killed bulk DM campaigns. Reply rates tank the moment the recipient smells automation.
The winning model is Claude drafts, you approve, you send. The human stays on the send button.
Does this work for cold outreach?
Not for the first message. Claude drafts to strangers are too smooth, and cold prospects can tell.
For cold outreach I write the first message myself. Claude handles the follow-ups once there is thread history.
What happens if LinkedIn changes their API?
Kondo handles the LinkedIn side. If LinkedIn changes anything, Kondo updates. You do not touch code. That is the whole reason to use Kondo instead of scraping yourself.
How do I write the voice file?
Open a blank markdown file. Write three sections.
Words I love. Words I hate. Banned phrases.
20 real sentences I have said or written. Three short paragraphs that sound exactly like me.
Keep it under 2 pages. Add to it every time Claude drafts something that does not sound like you and tell Claude what it missed.
Is Kondo the only way to do this?
No, but it is the cleanest. You could use LinkedIn's own export, or a Chrome extension, or paste threads into Claude by hand. All of those break or eat more of your day than the tool was supposed to save.
Kondo charges money in exchange for one less thing to worry about. That trade is worth it when your time is the bottleneck.
What if Claude's drafts sound wrong?
That is a voice file problem, not a Claude problem. Open your voice file. Add 10 more example sentences. Add the specific phrases Claude keeps using that you would never say. Re-run the draft prompt. Quality jumps.
If you skip the voice file, the drafts will always sound like a generic LinkedIn ghostwriter.
How many hours a week does this save?
I save two hours every weekday. Ten hours a week. Before this, I was answering DMs after dinner because morning time was burned on admin. Now my DMs are done by 7am.
Your number will be different. The ratio stays the same. Admin time drops to near zero. Writing time stays the same. Total time drops by around 75%.
Will prospects know I used AI?
If you keep the approval loop, no. Every message is something you wrote or rewrote. Claude is drafting, you are sending.
If you skip the approval loop and send Claude's raw drafts, prospects will know within two messages. They can smell it.
Does this work with other tools besides Kondo and Claude?
The pattern works with any inbox tool that has an API and any LLM that can read files. I picked Kondo and Claude because they are what I use. If you already have Linear or Airtable or Notion wired to your LinkedIn, adapt the prompts to your setup.
The only non-negotiable is the human-in-the-loop approval. Everything else is a tool swap.
Want the Full Prompt Library
Reply to my newsletter at sucana.ai/newsletter and I will send you the folder. Seven prompts, one voice file template, ready to paste.
