Prompt & Done · AGENT

The setup: make the reels with Claude Code, then let posting and replies run on their own

Start with the free kit below to make your first reel. Then the two things that run without me, and the files that keep Claude on track.

Numbers checked 10 Nov 2026: 5 channels, a post every morning on each (Nov 11–17), 141 posts scheduled, 65 Replybeam automations running, 405,200 views over 463 tracked posts (Instagram + YouTube + Facebook).

Download this first: REELKIT, the kit to make the reels

A free, open-source (MIT) starter kit: Claude Code researches and writes the script, a local AI voice reads it, Whisper checks every word, and a 3D scene renders frame by frame into a finished vertical video.

16 automated checks run on the final file, then you watch it and decide. The kit never posts.

Setup, the first prompt and the full stack: promptanddone.com/reelkit. Claude Code itself needs a paid Claude plan.

Replybeam: the comment replies and DMs

When someone comments your keyword, Replybeam replies to them publicly and sends your DM with the link. Automatically, in your own words: every reply and DM is written by you, never generated.

Free plan: $0 a month, no credit card. 2 running automations, 1,000 DMs a month, all four triggers (post comments, story replies, story mentions, DM keywords), link tracking and analytics, and the missed-comments and activity log. DMs on the free plan end with “Sent via Replybeam.app”. Pro is $9 a month (or $90 a year): unlimited running automations, no monthly DM limit and no “Sent via Replybeam.app” line.

You connect through Instagram's own login, no password. Not affiliated with Instagram or Meta.

What runs on its own once a reel is made

What isn't automatic: I approve every reel, and I say yes before each automation goes live. Comments without the keyword are not answered automatically.

1 Claude Code + one repo per channel

Each channel is its own folder (a git repo) that Claude Code works in. The channel's voice, look, rules and audience live there, so Claude never mixes two brands.

The real prompt that started one reel was a single sentence: "Make me one mind blowing reel, insane quality people would think they spent millions". Claude Code then made 106 tool calls over 58 minutes (research, script, voice, animation, checks) with no other typed prompt.

2 A CLAUDE.md that is the pipeline

Claude Code reads CLAUDE.md at the start of every session, so "make a reel" always runs the same way. Paste this skeleton into your repo and replace the <...> parts (or start with Claude Code's /init and add these sections).

# <Channel name>: how a reel gets made

Read this file and every file in rules/ at the start of each session.
"Make a reel" always runs the steps below, in order. Don't skip a gate.

## Rules
- When I correct you, save the correction as a new file in rules/
  (the rule, Why, How to apply) and follow it from then on.
- Every claim and number in a script needs a source in sources.md.
- Never post, schedule or send anything without my OK for that reel.

## Pipeline
1. Research: pick a topic people already watch; write reels/<slug>/sources.md.
2. Script: one spoken paragraph; each claim points to a row in sources.md.
3. Voice: text to speech, then speech to text; keep only takes that
   match the script word for word.
4. Scenes: built in code and rendered frame by frame.
5. Captions: timed word by word from the voice.
6. QA gate: run the checks on the finished file; fix and re-run until
   every line says PASS.
7. Package: caption, title, cover. Then stop and show me.
8. After my OK: schedule it in the next free slot and prepare the
   keyword reply + DM for my yes.

## QA checks (all must PASS)
- length is inside <your range>
- loudness about -14 LUFS
- first frame is not blank
- captions match the speech
- speech to text equals the script

3 A rules file it writes to itself

When you correct it, have Claude save the correction as one small file in rules/ that it reads every session. The template:

# <short rule name>

<The rule in one sentence: what to always do, or never do.>

**Why:** <what happened that made this a rule: the date, and your own words
if you said something when you corrected it.>

**How to apply:** <when it kicks in and what to do instead, concrete enough
that a fresh session can follow it without asking you.>

The real one from the reel, written after I rejected three copycat reels:

After SF Milk hit 12k, I built 3 more "Who leaves ONE…" copies.
The owner rejected and deleted them (8 Nov 2026): "why does it have
to be same thing over and over… people will leave".

How to apply: follow a winning format with at most 1–2 close
follow-ups. After that, each new reel needs a different premise,
comedy mechanic and staging.

4 A QA gate before anything ships

A script checks the finished video, and a reel only moves on when every line says PASS. The real check from one Stickfigure reel (8 of 8 passed):

PASS  duration  18.09 s
PASS  loudness  -14.2 LUFS, true peak -3.7 dBTP
PASS  frame 1   moving
PASS  captions  all script words in order
PASS  sync      95th pct 17 ms
PASS  asr       WER 0.000 vs the script
PASS  post      title 40 chars
PASS  cover     text inside the grid crop

Good first checks: length, loudness (about -14 LUFS), the first frame isn't blank, captions match the speech, and a speech-to-text pass equals your script. REELKIT ships with 16 of them.

5 A scheduler with an API

Finished reels go into a scheduling tool through its API, so Claude can fill the calendar itself (we keep five morning slots, one per channel). Any scheduler with an API or an MCP server works; give Claude the key in an environment variable, never in the repo.

6 Setting up the Replybeam automation from Claude Code

On Replybeam's Pro plan, its AI connector (MCP) lets Claude create the keyword automation for each reel. The DM text and replies are still the ones you wrote. A real call (5 Nov 2026, shortened):

create_automation
  trigger: post_comments   keywords: SEED, SEEDS
  dm_message: "Hey {username}! Here's the world from the reel…"
  public_replies: "Sent! Check your DMs", …
→ "It's running: matching comments and messages get answered from now on."

On the free plan you set up the same automation in the app yourself. Either way, keep a human yes before an automation goes live: once it's on, real people get the message.

7 What it doesn't do

It doesn't decide what's good: I still judge every reel and say no when it's wrong (that's where the rules come from). The 405,200 views are spread over 463 posts, not one viral reel.