a free guide by
Luka@omgluka.human
21 real notesfrom 3 reels

My AI agent picks the take, builds the timeline, grades the color and designs the titles. It cannot tell me which version is better. That part is notes. Here are the 21 I gave on three reels, word for word, what each one changed, and the rule it became.

the order of the work
It does most of the edit
01It does most of the editthe take, the timeline, the grade
It can't tell what's good
02It can't tell what's goodmy notes, today
Taste is the job
03Taste is the jobthe part that stays mine
The reel this guide comes from, in three beats: what my agent does, what it cannot do (two of my real notes), and what is left for me.
sound familiar?
  • The first cut looks fine, but it does not look like yours.
  • You give the same note on every video.
  • “Make it better” and nothing really changes.

A note you give twice is a note you did not save.

before you start

What you need

An agent that edits, and something to react to.

  • An AI editorClaude Code with editing skills. Mine are in the 3 editing skills I kept.
  • A first versionLet the agent make one before you say anything. Notes are reactions, not instructions up front.
  • A place for notesA chat works. A feedback log works better. In DaVinci Resolve, markers on the timeline work best.
step 01

How to read a note

Every note below has three parts. The rule is the one that matters: it is what I never have to say again.

  1. 01
    The noteWhat I typed, word for word. Short, lowercase, no polish.
  2. 02
    What changedWhat the agent did with it in the next version.
  3. 03
    The ruleWhat the note became, so I never have to give it twice.
one note, before and after
A reel hook before and after a note: CLAUDE big on the left, EDITING SKILL big on the right
“claude is too big, highlight should be on editing skills”, on the first version of a reel. Before on the left, after on the right.
notes on · 3

The first second is yours to design.

The agent will put a title on the hook. What the eye lands on first is your call.

“hook title bolder, multiple lines, mini icon of dvr, designer-y”

One thin line became a four line stack, with the real DaVinci Resolve app icon next to its name.

the ruleA hook title is a designed lockup, not a caption. Real app icons make it look designed.

“hook title can be better designed, claude is too big, highlight should be on editing skills”

The brand name shrank into a small chip with its icon. The gold highlight moved to the two words the video is about.

the ruleHighlight the topic the viewer cares about, not the brand. The tool goes in a small chip.

“improve hook title, center it, designery”

The title was centered above my face: a small line on top, the gold word, the old tool struck through.

the ruleCenter the title when the face is below it. Never put text on the face.

what they became
Three reel hooks side by side
The three hooks as they shipped: the real DaVinci Resolve icon in a stacked lockup, the highlight on the two words the video is about, and the title centered above my face with the old tool struck through.
notes on · 3

If they cannot read it, it is not there.

Agents love small text and dark boxes. You are the one holding the phone.

“white text is a bit too small now”

The white hook lines went from 62 to 88 pixels, stacked on two lines, with a stronger shade behind them.

the ruleHook text is at least about 88 pixels tall on a 1080 wide frame. Check it on a phone.

“comment motion radial black should be blurred and better designed”

“line still has rough cut”

The black blob behind the text became a frosted glass card. A shade with a visible edge became a soft blurred one.

the ruleNo black blobs and no hard edges behind text. Use frosted glass or a soft blur, and keep cards off the head.

“improve colors and layout”

Every title now sits in the same zone. Backgrounds got darker and less saturated, so the panels in front stand out.

the ruleOne place for titles. Backgrounds step back, the subject pops.

what they became
A frosted call to action card and a title in a fixed zone over a dark background
The black blob became frosted glass on the call to action, and every title sits in one zone over a darker background.
notes on · 4

The agent picks footage. You pick what it means.

It can find a clip of you typing. It does not know which clip feels right.

“use better footage for each scene”

Paused screenshots were replaced by the raw clip and the finished clip playing together, frame for frame.

the ruleShow real motion and real change. Never a static screenshot.

“change the footage you used for first skill demo”

A messy phone selfie was replaced by a take from my main camera at the desk.

the ruleDemos use your best looking camera, not a raw phone clip. Check the take has real speech first.

“don't use the part where I do the hand gesture to call them to see”

The typing shot moved to a part of the clip with no gesture in it.

the ruleA gesture means something. Do not use it under the wrong line.

“better intro video, my fonts”

The hook moved to a real reaction shot, and every title switched to my own typefaces.

the ruleThe hook face is a real reaction. The type is always yours.

what they became
A before and after of a shot playing together, and a desk take from the main camera
Raw and finished playing together instead of a paused screenshot, and the demo take from my main camera at the desk.
notes on · 3

How it moves, and where it stops.

Two things an agent will not feel on its own: a title that lands well, and a shot that runs too long.

“improve letter animation, blur fade in letter by letter”

Every title now fades in one letter at a time, from blurred to sharp.

the ruleTitles animate letter by letter. It is the default, not a special effect.

“last video should end earlier”

The last shot went from 4.9 seconds to 2.8, and now ends while the gesture is still held.

the ruleEnd on the payoff. No tail.

“me at a computer and cutout screen for approving”

The ending became me at the computer: the screen lifts out as a card, Approve gets clicked, I give a thumbs up.

the ruleTie the payoff to a person on camera, not to a screen alone.

what they became
The review card with Approve clicked, and the last shot ending on a thumbs up
The payoff on a person: me at the computer, the review card, Approve. And the last shot now stops on the thumbs up.
notes on · 3

Only what is true. Only what is needed.

An agent will happily add a step that sounds right. You are the one who knows if it is.

“no need for the setup you mention”

My guide listed install steps that a new app version had made unnecessary. It was rewritten to one update and one sentence.

the ruleNever add a setup step that is not needed.

“under my whole prompt add mini cards with names of skills”

Small cards under the prompt now name the skills that did the work.

the ruleShow what actually did the work.

“remove code from the screen, explain AI is very good at writing code”

A wall of code became three plain step cards and one takeaway line.

the ruleExplain the idea. Do not show raw code.

what they became
A prompt card with skill cards under it, and three plain step cards
Cards under the prompt that name the skills that did the work, and three plain cards instead of a wall of code.
notes on · 2

It has to look like the source.

If the export looks different from your raw clip, something in the middle changed it.

“video quality broken, color is different”

The phone clip was HDR. It was converted to standard color the right way, and the grade and the vignette were made lighter.

the ruleCompare the color with your raw clip. HDR phone footage gets converted before anything else.

“final video should be 4k if that's best output”

The master is now true 4K, built from 4K sources. Instagram gets a smaller copy made from it.

the ruleKeep a 4K master. Downscale for the platform, never the other way around.

notes on · 2

You approve the cut. Then it renders.

The agent is fast enough to finish the wrong thing. Decide where it has to stop and ask.

“never render final video, always show in editor for me to confirm”

The agent now stops at a live preview and waits. No final file exists until I say yes.

the ruleNo final render before you approve the cut in the editor.

“perfect, remember this”

I said it when a hook design finally landed. That design became the standard for every reel after it.

the ruleWhen a note lands, say so. “Remember this” turns one good result into the default.

where it stops now
HyperFrames Studio playing a vertical reel with its timeline
My agent stops at a live preview in HyperFrames Studio and waits. No final file until I say yes.
step 09

Notes an AI can act on

You do not need to sound like a director. You need to be specific.

  1. 01
    Say what is wrong and what you want“claude is too big, highlight should be on editing skills” has both halves in one line.
  2. 02
    One note per problemThree problems are three notes. A paragraph gets half fixed.
  3. 03
    Point at the momentThe second, the scene number, or a screenshot. “Especially scene 3” is enough.
  4. 04
    Say why when it is taste“Better designed” makes it guess. The reason tells it what better means.
say to Claude

At [second], [what is wrong]. I want [what you want instead], because [why].

A shape to copy until it is a habit.
point at the moment
A timeline with red markers, and three notes with their timecodes
In DaVinci Resolve my notes sit on the timeline as red markers, each on its frame, with the note typed in. The whole method is in my Resolve guide.
step 10

Turn your notes into a skill

A note you give twice is a note you did not save. Keep a feedback log, then make the agent read it before every edit.

  1. 01
    Walk through one editLet the agent make a first version. Give your notes like you would give an editor.
  2. 02
    Log every noteThree columns: the note, what changed, the rule.
  3. 03
    Save the rules as a skillFrom then on the agent starts from your taste, not from zero.
say to Claude

Add that note to my feedback log: what I said, what you changed, and the rule it becomes.

After each note.
say to Claude

Save these rules as a skill and read it before every edit.

When the edit is right.
what goes in it
A card listing what goes into an edit skill
My edit skill, as one of my reels lists it: how I pace a cut, what counts as dead air, my jump cuts and zooms, my colour look.
step 11

The feedback log

Three columns. One row per note. The first row is one of mine.

Note
What changed
The rule
“white text is a bit too small now”
Hook lines 62 to 88 pixels, on two lines
Hook text at least about 88 pixels
say to Claude

Start a feedback log for my edits with three columns: Note, What changed, The rule. Add a row every time I give you a note.

if something breaks

Quick fixes

The new version is not what I meant

Before it changes anything, ask it to say back what your note means. If its reading is off, you fix one sentence instead of a whole version.

It fixed one thing and broke another

Give one note per problem and look after each round. Ask it to list everything it changed since the last version, so you can spot what moved.

It keeps making the same mistake

The note never became a rule. Add it to your feedback log and your skill (steps 10 and 11), and ask it to read the skill before every edit.

I can tell it is off, but not why

Put a reel you love next to yours and ask: “What does this one do in the first second that mine does not?” Then turn the answer into a note.

My note was wrong

That happens. Ask it to check before it changes: measure the clip, compare the frames, and tell you when a note does not match what it sees.

bonus · prompts I use when I give notes
say to Claude

Before you change anything, tell me in one line what you think my note means.

say to Claude

Show me two versions of the hook and wait for my pick.

say to Claude

List everything you changed since the last version.

say to Claude

Read my feedback log before you start this edit.

An independent guide by Luka (@omgluka), made at Engram. Every note here was given on a real reel. Claude is made by Anthropic and is not affiliated with this guide.

next · the AI agents course

Let agents run the rest of your work.

My AI agents course: build and train custom agents that run the processes in your business, on one real job, step by step.

4 acts · 28 episodes · hands-on, for your business · from $79 a month

join the waitlist

Opening soon. One email when it opens, nothing else. More guides like this one on engram.media/guides and on Instagram, @omgluka.

Rather have it built for your team? Get a written two-option proposal.