start here · read me first

One week to studio speed.No AI experience needed.

Welcome to Convey, Rafael. You already know how to edit. This course teaches the other half: how to direct a team of AI agents that do most of the production work, so your time goes where your talent is. By the end of the week you will understand the whole company, run our editor like it's your own, and know exactly how to be great in this seat.

Rule of the course: do the "Do this now" boxes. Reading alone teaches 20 percent. Doing teaches the rest.

How the week works

Day	What you master	Time
Day 1	What Convey is, and exactly what your job is	~1.5 hrs
Day 2	How AI actually works, in plain English, and the theory of being great with it	~2 hrs
Day 3	Prompt engineering: briefing agents so the output is right the first time	~2 hrs
Day 4	Convey Studio, our editor, top to bottom	~2.5 hrs
Day 5	The full agent team and how to orchestrate them	~2 hrs
Day 6	Teaching the system: context building, agent training, finding tools	~2 hrs
Day 7	The complete loop, your first real edit, and what "the best" looks like	~2.5 hrs

Each day ends with a checklist. Tick every box (they save automatically), press Mark day complete, and move on. Days build on each other, so go in order.

What you need before Day 1

Your machine set up (the one-line installer Dougie texted you, or WINDOWS-SETUP.md in your claude-setup repo)
Your two GitHub invites accepted (the brain + claude-setup)
Your onboarding email saved (subject starts with "SAVE THIS ONE")
The Studio password from Dougie (you get it on your setup call)

Missing any of these? Message Dougie before starting. Everything in this course assumes they're done.

Ready?

day 1 · the company and your seat

What Convey does,and exactly what you do inside it.

1.1 The business, in three sentences

Convey installs and runs the entire growth engine for e-commerce brands: their website, their email, their ads, and the content feeding all of it. Brands pay us a one-time install and then a monthly retainer to keep operating it. We can charge less than agencies and deliver more because AI agents do most of the production work, supervised by a small team of humans who are excellent at directing them.

Right now the live brands are Square One Coffee (specialty coffee, voice: "the kindest troublemakers in coffee") and Sweet Addison's (protein cookies and brownies). More are signing. Every brand needs a constant stream of edited video and static ads. That stream is your job.

1.2 The one mental model to hold all week

You are the creative director of a small AI studio. Research agents find who to talk to and what's winning. Maker agents build to about 70 percent. You direct, review everything, and finish the last stretch where your talent lives.

Stage	Who does it	What happens
Research	Delphi + Intel Ann (agents)	Who the audience is; which competitor ads are proven winners
Make	Eddie (agent) in Convey Studio	Selects moments, cuts, captions, music, color grade, first render
Judge	The judge (automatic)	Scores the render against real winning content before you see it
Review	You	Your eyes and ears on every piece. Fix by prompt or by hand
Finish	You, in Premiere / After Effects	The last stretch that needs a real editor
Ship	You approve	Final files to the brand's Drive folder. Nothing ships without a human

1.3 The four rules that never bend

rule 1

A human approves every piece

Even a perfect one. You produce; a human releases. Nothing goes live on its own, ever.

rule 2

Never let AI redraw a product label

AI turns label text into gibberish. Always composite the real product photo, then read the label letter by letter before approving.

rule 3

Each brand stands alone

Load the brand's style preset and its never-say list before touching a frame. Never recycle another brand's look.

rule 4

Never fabricate, never cut harsh

No invented numbers, reviews, or claims. No mid-word cuts, no audible seams; verify by listening. House style: no em-dashes, no stock emojis.

1.4 Where everything lives

Place	What it is
The brain	Our GitHub repository (conveygrowth/convey-brain). Every brand's context, every agent, every playbook. You pull it at the start of every session
Convey Studio	The video editor we built. Hosted; opens in your browser. Day 4 is all about it
Google Drive	The content hub, one folder per brand. All raw footage in; all approved finals back in. If it's not in the Drive, it doesn't exist
Claude Code	The chat panel in your VS Code where you brief the agents in plain English
The operating manual	brain.conveygrowth.com → manual, the standing reference for everything

Do this now (20 min)

Open your cloned brain folder in VS Code (code %USERPROFILE%\convey-brain or File → Open Folder).
Read onboarding/editor/EDITOR-START-HERE.md top to bottom.
Browse clients/ and open Sweet Addison's folder. Just look around; this is the kind of context agents read.
Open the two brand Drive folders from your email invites and skim what's in Content.

1.5 Check yourself

A brand asks why your edits are cheap compared to their old agency. What's the honest answer?

AI agents do about 70 percent of the production work, so human time (the expensive part) goes only where it's needed: direction, review, and finishing. Same or better output, less human hours per piece.

Eddie produces a perfect cut. It scored high with the judge. Can it go straight to the client?

No. Rule 1: a human approves every piece, no exceptions. You watch it, full, with sound, before anything ships.

Where do finished, approved edits go?

The brand's Google Drive folder. The Drive is the single source of truth for content.

1.6 Day 1 checklist

I can explain what Convey sells in one sentence I can name the six stages of the mental model, and which ones are mine I know the four unbendable rules without looking I opened the brain in VS Code and read EDITOR-START-HERE I opened both brands' Drive folders

Ticks save automatically in this browser

All boxes ticked?

day 2 · how AI actually works

AI, in plain English.And the theory of being great with it.

2.1 What a model like Claude actually is

Claude read a huge portion of everything humans have written, and learned to predict what text should come next. That simple trick, done at enormous scale, produces something that can reason, write, plan, and use tools. Two consequences matter for your job:

consequence 1

It only knows what's in front of it

Claude has broad general knowledge, but it knows nothing about Square One's voice or Sweet Addison's allergen rules unless that text is in its "context": the files and messages loaded into the conversation. Feeding it the right context is most of the skill.

consequence 2

It's confident even when wrong

A model predicts plausible text, and "plausible" sometimes means "made up." It will say a task is done when it isn't. That's not lying, it's the failure mode of the technology, and it's why verification is the human's half of the partnership.

2.2 The vocabulary you actually need

Word	Plain English
Context	Everything the model can currently "see": your messages, files it read, its instructions. The single biggest lever on quality
Prompt	What you type. A brief. Day 3 makes you great at these
Agent	Claude plus a job description (a markdown file) plus tools. Eddie is Claude wearing an editor's job description
CLAUDE.md	The rulebook file Claude reads automatically at the start of every session. Yours is installed already
Skill	A packaged workflow Claude can load: clip selection, captioning, and so on. Your ten are installed already
Hallucination	Confident, plausible, wrong. The reason you always verify
Token	The chunks of text models read and write. Only matters to you as "context is finite, keep it relevant"

2.3 The three laws of being great with AI

This is how Dougie and Dale get 10x more out of the same tools as everyone else. All three laws, every day:

law 1 · context beats cleverness

A mediocre prompt with the right context beats a genius prompt with none. Before asking for work, ask: does Claude have what it needs to know? The brand's files, the reference reel, the actual footage, the never-say list. At Convey most of that context is pre-built (the brain, the presets, EDITOR-CONTEXT.md), your job is to point at it: "Read the SQ1 brand file first, then..."

law 2 · trust the work, verify the claim

Let agents run without babysitting every step, that's what makes you fast. But never accept "done" as evidence. Watch the render. Listen for seams. Read the label letter by letter. The verification takes minutes; a bad piece reaching a client costs trust you can't buy back.

law 3 · never fix the same thing twice

When an agent gets something wrong, you can fix the output (helps once) or fix the instructions (helps forever). Every correction belongs in the agent's file or the brand's preset. This is Day 6's whole topic, and it's the difference between using the system and compounding it.

Dale's phrasing: "The AI is a brilliant intern with amnesia. Write everything down where it will re-read it."

2.4 Talking to Claude: the beginner mistakes

Mistake	Do instead
Asking tiny questions one at a time	Give the whole job with all constraints up front. Claude handles big, multi-step asks
Accepting the first output	Iterate in plain English: "warmer", "faster intro", "captions lower". Small nudges are cheap and fast
Arguing with a bad result	Start a fresh chat with a better brief. A conversation gone sideways stays sideways
Being polite instead of specific	"Make it pop" is noise. "Bigger hook text, first cut at 0.8s, brand coral" is signal
Treating it like search	It's a collaborator. Give it a role, context, a task, and what "done" looks like

Do this now (30 min)

Open Claude Code in VS Code (in the brain folder). Say: "Read onboarding/editor/EDITOR-CONTEXT.md and hold it as standing context. Then explain my role back to me in five bullets." Notice how the context file changes what it knows.
Then ask: "What are Sweet Addison's never-say rules for content?" It will find and read the brand's files itself. That's law 1 working.
Deliberately ask something it can't know: "What did Dougie say to me yesterday?" Watch how a model behaves at the edge of its knowledge. Remember that behavior.

2.5 Check yourself

Your edit brief keeps producing off-brand captions. Which law applies first?

Law 1: check the context. Is the brand preset loaded and named in the brief? If the context is right and it still misses, law 3: write the correction into the preset so it never happens again.

An agent reports "render complete, judge passed, pushed to review." What do you actually know?

Only that the agent claims it. Law 2: open the render, watch it with sound, then you know. Claims are not evidence.

2.6 Day 2 checklist

I can explain "context" and why it beats clever prompts I can say the three laws from memory I loaded EDITOR-CONTEXT.md into a session and saw the difference I asked Claude for Sweet Addison's content rules and read its answer I know the five beginner mistakes and their fixes

All boxes ticked?

day 3 · prompt engineering

Briefing like a director.Right the first time, at volume.

3.1 The anatomy of a great brief

A prompt is a creative brief. The same thing you'd give a human editor, except this editor reads instantly, never gets bored, and takes you completely literally. Every great brief has five parts. Memorize these five; they are the skill.

Part	What it does	Example
1 · Agent	Who's doing the work	"Eddie:"
2 · Brand	Pulls the right voice, preset, guardrails	"SQ1 Coffee"
3 · Source	The exact footage or material	"today's founder clip in SQ1's Content folder"
4 · Preset + lane	The look and the purpose	"ugc-cozy preset, ad lane"
5 · Done looks like	Length, hook, platform, the finish line	"under 30s, hook on the 'first cup' line, push to review when the judge passes"

the five parts, assembled · copy and adaptEddie: SQ1 Coffee, TikTok. Source: the newest raw clip in SQ1's Content folder on the Drive. Use the ugc-cozy preset, ad lane. One cut under 30 seconds, hook on the "first cup" line in the first second, captions per preset, push to review when the judge passes.

3.2 Vague in, slop out

the same request, done badly"can you make a cool tiktok from the coffee footage?"
why it failsNo brand context loaded. No source named (which footage?). No preset (defaults look generic). No lane. No definition of done, so the agent guesses all five, and every guess is a coin flip you'll pay for in revisions.

The math that makes this matter: a tight brief costs you 60 extra seconds. Five rounds of "no, not like that" costs 30 minutes, and worse, it costs your attention. At the volume this job runs, briefs are where your day is won or lost.

3.3 The patterns that level you up

pattern · context first, then the ask

"Read the SQ1 brand file and the ugc-cozy preset. Then..." One sentence of pointing saves ten sentences of describing. The brain exists so you can point instead of explain.

pattern · show, don't describe

Taste is hard to type. Instead: "Match the pacing of this reference reel [link]" or "like version 3 but with the v1 hook." Agents are excellent at matching examples and mediocre at decoding adjectives.

pattern · ask it to ask

On anything complex, end with: "Ask me up to five clarifying questions before you start." Two minutes of questions kills an hour of wrong work. This is the single most underused move in prompting.

pattern · batch, then review in one pass

"Give me 5 hook variations for this cut, one line each, then hold." Review all five at once, pick one, continue. Batching turns you from a lever-puller into a director.

pattern · make it verify itself

"Before you say done: confirm the judge score, confirm no cut lands mid-word, list anything you're unsure about." You still verify with your own eyes (law 2), but self-checks catch the cheap misses first.

3.4 Iterating without starting over

First output is a draft, not a verdict. Nudge in plain English, one dimension at a time:

nudges that work · all real"faster intro, first three cuts under a second each"
"captions one band lower, they're covering her eyes"
"the music swell fights the voiceover at 0:14, duck it harder"
"keep everything, swap the hook line for variation 3"
"this feels like an ad, make it feel like a friend's story"

↳ when a chat has gone sideways twice, stop nudging. Fresh chat, better brief. Ten seconds, clean slate.

3.5 Exercises

Do this now (45 min)

Rewrite drill. Take the bad prompt from 3.2 and rewrite it with all five parts, from memory. Then check against 3.1.
Live run. In Claude Code: "Read EDITOR-CONTEXT.md. Then: which raw clips currently sit in Sweet Addison's Content folder on the Drive? List filenames and lengths, newest first." A real brief with a verifiable answer, go check the Drive yourself and confirm it told the truth. That's law 2 as a habit.
Ask-it-to-ask. Brief a first cut of any SA clip, ending with "ask me up to five clarifying questions first." Answer them. Notice how much sharper its plan gets.

3.6 Check yourself

Name the five parts without scrolling up.

Agent · Brand · Source · Preset + lane · What done looks like.

You can't describe the vibe you want. What's the move?

Show, don't describe: hand it a reference (a reel link, an earlier version) and say "match this." Or have Studio learn the reference as a preset (Day 4).

3.7 Day 3 checklist

I wrote a five-part brief from memory I ran the Drive-listing brief and verified the answer against the real Drive I used "ask me up to five questions" on a real brief I know all five leveling patterns I know when to nudge and when to start a fresh chat

All boxes ticked?

day 4 · convey studio

The editor we built.A timeline your agent drives.

4.1 What Studio is

Convey Studio is a real video editor, version 3, that an agent operates while you watch the timeline move live in your browser. You hand it raw footage; it transcribes, cuts dead air, writes brand captions, adds music and graphics, grades the color, renders, and scores its own work before you ever see it. Nothing installs on your machine; you sign in once with the password from Dougie.

Do this first (60-90 min, the heart of the day)

Watch the full official walkthrough. Thirty-five short chapters, about 29 minutes total, every part of the screen named and explained, with an annotated map you can hover:

Open the Studio tutorial →

4.2 The five ideas that make Studio different

idea 1 · the transcript is the edit

Cutting happens by words. Click a word to jump there, select a range, Remove. The removed words stay on screen struck through, so you always see what left the cut. Cutting speech by reading beats scrubbing waveforms, every time.

idea 2 · nothing is ever deleted

Every cut becomes a "ghost": still in the timeline, disabled, carrying a note of who cut it and why. Hover a ghost scar and the card shows the reason and where it used to sit. Any decision can be undone, in Studio or later in Premiere.

idea 3 · every change carries a written reason

The agent must justify each edit in plain English, and that reason follows the cut into the Premiere export. You're never guessing why something happened; the edit explains itself.

idea 4 · presets outrank everyone's taste

In the Style References panel, paste a reel the brand loves and press learn. Studio measures it (grade, cut rhythm, caption style, music sync) and writes a rulebook, the preset, that outranks Studio's defaults and the agent's own taste. A cut reading "no preset set" is a flag to fix, not a style choice. Each cut also carries a lane: ad, organic, viral, or client.

idea 5 · the judge scores before you review

Every render is graded against a rubric plus a benchmark built from real winning reels in the niche. The judge chip shows the score on the version you have open. Fail a dimension, the agent fixes and re-renders. Your review starts where the machine's standards end.

4.3 The screen, at a glance

Region	What it's for
Brand rail (far left)	One tile per brand; click to switch. A brand joins Studio by handing over its Drive folder
Left dock	Six panels, two open at once: Bin, Script (the transcript), Graphics, Audio, Feed (the decision log), Style References
Monitor (center)	Renders captions and graphics with the same components the final render uses. What you see is what ships
Safe zones (right of monitor)	TikTok, Reels, and Shorts drawn as those apps actually cover the screen, plus the face band captions avoid
Header chips	Version, Edit History, lane, preset, judge score, server status, Export (MP4, Premiere ghost XML, SRT)
Edit History	Every version, its score, and what it did in plain English, grouped under the chat prompt that caused it

4.4 The Premiere handoff

When a cut needs your hand, export the Premiere ghost XML. It opens as a normal timeline where every cut Studio made still exists, disabled, annotated. You finish the way you always have, except the whole edit history came with you. Your rule of thumb: Studio to 70, you to 100. If you're rebuilding from zero in Premiere, something upstream (the brief or the preset) needs fixing first.

Do this now (30 min, after the tutorial)

Open Studio, pick Square One on the brand rail, open any existing cut.
Find one ghost cut in the timeline. Hover it, read the reason.
Open Edit History. Read what version 1 did versus the latest.
Open Style References and look at ugc-cozy's color ramp. That's a measured look, not a vibe.
In the transcript, click any word and watch the playhead jump. Don't cut anything yet; tomorrow you drive.

4.5 Check yourself

A client hates a cut you shipped last week. Can you show them exactly what was removed and why?

Yes, twice over: the ghost cuts in the timeline each carry a written reason, and Edit History shows every version with what it did in plain English. Nothing was deleted, so anything can be restored.

The header reads "no preset set." What do you do?

Stop and set the brand's preset (or learn one from a reference) before serious work. Defaults produce generic output, and generic is off-brand for every brand.

4.6 Day 4 checklist

I watched the full 35-chapter tutorial I can explain ghosts, reasons, presets, lanes, and the judge in my own words I opened a real cut, read a ghost's reason, and browsed Edit History I looked at a learned preset's color ramp in Style References I know when to export the ghost XML versus asking for another Studio pass

All boxes ticked?

day 5 · the agent team

Your team,and how to conduct it.

5.1 The roster

Every agent is Claude wearing a different job description, a plain-English markdown file in the brain's agents/ folder. You hire one by naming it in your brief. These eight are your working team:

Agent	Their job	Call them when
Eddie	Lead editor; runs the whole Studio pipeline	Any video edit, start to finish
Intel Ann	Competitor ad intelligence across Meta, Google, TikTok. Long-running ads with many variants = proven winners	Before editing, so you build on proof, not guesses
Delphi	Audience research: who the customer is, what they care about	The angle is unclear, or a new brand lands
Lumen	Hyper-real generated imagery for b-roll and statics	You need a shot that doesn't exist in the footage
Mr. Stat	Static image ads for Meta and Google	Image ads. His law: real product photo composited, never AI-redrawn labels
Designer	Layout, typography, brand-kit enforcement	A static or card needs real design polish
Portia	Compliance: claims, regulated products, platform rules	Anything claim-adjacent, especially Sweet Addison's protein products
Roast	The professional skeptic; stress-tests work	Before you call something important "done"

5.2 The three workflows you'll run daily

workflow A · edit from a brief

Brief Eddie (five parts) → watch the Studio timeline → judge passes → your review with sound → nudge or ghost-XML to Premiere → approve → final to Drive. This is 80 percent of your day.

workflow B · find the angle first

No brief? Build one: Delphi (who is this for) → Intel Ann (what's proven in this niche right now) → pick the strongest angle → workflow A. You're now sourcing your own briefs, which is what senior looks like.

workflow C · statics

Intel Ann for winning formats → Mr. Stat builds with the real product photo composited → the label check, letter by letter → Portia if any claim appears → approve → Drive. Statics finish in Photoshop or Figma, never Premiere.

5.3 Compliance, the part that can hurt a brand

Sweet Addison's sells protein cookies. The words "builds muscle," "boosts recovery," "energy," or "healthy" as a badge can trigger regulatory trouble for the whole shop. Food language only. When any health-adjacent claim appears in a hook or caption, Portia checks it before it ships. When unsure, don't ship, ask.

Same discipline on facts: Square One ships "roasted and shipped within 48 business hours," never "the day it ships." Never invent an award, a stat, or a review. Uncleared means leave it out.

5.4 Orchestration: running more than one thing

Agents don't get tired and don't need you watching. The senior move is parallel direction: brief Eddie on cut one, and while Studio works, have Intel Ann research tomorrow's angle, then come back to review. Two rules keep this safe:

One brief, one finish line. Every parallel thread has a clear "done" so nothing dangles half-finished.
Review is serial. Work runs in parallel; your approval attention goes one piece at a time, fully.

Do this now (40 min)

Open agents/eddie.md and agents/intel-ann.md in the brain. Skim them. Notice they're just clear English; there's no magic.
Run workflow B for real: ask Delphi who Sweet Addison's core customer is, then ask Intel Ann what cookie and snack ads are currently proven winners. Save both answers; they're tomorrow's brief.
Ask Portia: "Review this caption for Sweet Addison's: 'the protein cookie that fuels your recovery.' Allowed?" Read why not, and what phrasing survives.

5.5 Check yourself

A hook says SQ1 coffee is "award-winning." You haven't seen that in the brand file. Ship it?

No. Never fabricate, and uncleared means leave it out. Check the brand's cleared facts; if it's not there, cut the claim or ask Dougie.

What's wrong with editing before running Intel Ann on a new brand?

You'd be guessing at what works in that niche. Ann surfaces formats already proven by competitors' long-running ads, so your first cut starts from evidence.

5.6 Day 5 checklist

I can name all eight agents and when to call each I ran Delphi and Intel Ann for real and saved the answers I tested Portia on a risky caption and understood the verdict I read two agent files and understand agents are just written instructions I know the two rules of running work in parallel

All boxes ticked?

day 6 · teaching the system

Make it smarter.This is the compounding part.

6.1 Why this day matters most

Everything before this made you effective. This day makes you more effective every week without working more hours, because at Convey the system itself is editable. The agents are text files. The presets are measured and re-measurable. The context is a repository you have write access to. When you improve any of them, the improvement is permanent, for everyone.

Law 3 from Day 2, now as a job duty: never fix the same thing twice.

6.2 Teaching an agent

Say Eddie keeps placing hook text where TikTok's UI covers it. You could re-prompt every time (fixing it forever, one edit at a time), or:

Open agents/eddie.md in the brain.
Find the right section (usually Rules or Gotchas) and add one plain-English line: "Hook text always sits inside the platform safe zone; on TikTok keep it above the caption band and below the top 8 percent."
Commit the change (Claude can do this for you: "commit this with a clear message").

Rules for teaching well: surgical edits (add or refine a line; never rewrite a working file), say why (agents follow reasons better than bare orders), and one lesson per edit so history stays readable.

6.3 Teaching a brand (presets and context)

New look? Style References → paste the reference reel → learn → name it. The measured preset now governs future edits. Better than any adjective you could type.
Recurring brand fact? ("SA's founder pronounces it Ah-di-son", "SQ1's bag always faces label-out") It belongs in the brand's folder in the brain, where every agent reads before working.
A win worth repeating? When a cut performs, note what made it work in the brand's context. Winners should echo.

6.4 Finding tools: how to hunt GitHub like Dougie and Dale

Part of your job is spotting tools and repositories that make our editing context stronger. The hunt method:

where to look

GitHub search with editor-shaped queries: "whisper subtitle", "silence removal video", "ffmpeg workflow", "remotion template", "auto captions". Sort by recently updated. Also: the "Used by" and "Dependents" tabs of tools we already run.

how to judge a repo in 90 seconds

Alive? Commits in the last few months, issues getting answered.
Adopted? Stars are a hint, real usage in issues/discussions is better.
Licensed? MIT, Apache, BSD mean usable. "No license" means look elsewhere.
Understandable? If the README can't explain it, integration will hurt.

what we already build on (study these)

openai/whisper (transcription under everything), remotion-dev/remotion (programmatic video, the toolkit's template), WyattBlue/auto-editor (the silence-cut engine behind auto-cut-timeline), ffmpeg (the renderer's foundation). Reading how these work makes you better at directing the skills built on them.

The proposal path: found something promising? Don't wire it in solo. Test it on scrap footage, write three sentences (what it does, what it would improve, license), and send to Dougie and Dale. If it's a yes, you'll usually build the skill wrapper with Claude, which is its own excellent training.

6.5 The feedback loop, weekly

Friday habit, 20 minutes: what did I correct more than once this week? Each repeat offender gets written into an agent, a preset, or the brand context. Five weeks of this and your agents feel psychic. They're not; you taught them.

Do this now (40 min)

Make one real teaching edit: add one Gotcha line to any agent file based on something you noticed this week, and commit it. Small is fine; real is the point.
Hunt one repo with the method above (start from "auto captions" or "silence removal"). Judge it with the 90-second checklist. Write the three-sentence proposal, even if the verdict is "not worth it", the muscle is the point.
Ask Claude: "Show me the last 10 commits to agents/ in this repo." Read how others' teaching edits look.

6.6 Check yourself

Eddie made the same caption mistake twice this week. What exactly do you do?

Open agents/eddie.md, add one surgical line in Rules/Gotchas stating the fix and why, commit it. Fix the instructions, not just the output.

A repo looks perfect but has no license file. Use it?

No. No license means no permission. Flag it to Dougie/Dale if it's genuinely special, but don't build on it.

6.7 Day 6 checklist

I made and committed one real teaching edit to an agent file I know the surgical-edit rules (add, don't rewrite; say why; one lesson per edit) I hunted and judged one GitHub repo with the 90-second checklist I wrote a three-sentence tool proposal I put the Friday 20-minute loop in my calendar

All boxes ticked?

day 7 · the full loop

Your first real edit,and what "the best" means here.

7.1 The graduation run

Today you run the entire loop on real footage, alone, start to finish. Everything you've learned, in order. Take your time; correctness beats speed on your first lap.

The graduation run (2 hrs)

Pull. In the brain: git pull. Always first.
Research. Use your saved Delphi + Intel Ann answers from Day 5 (or refresh them) and pick one angle for Sweet Addison's.
Brief. Five parts, "ask me up to five questions first." Answer them.
Watch. Studio timeline moves. Read at least two of the agent's reasons while it works.
Judge, then you. When it passes, watch the render fully, with sound. Nudge at least one thing (there's always one thing).
Finish. Export the ghost XML, open in Premiere, make one real human improvement, even small.
The gate. Label check if any product appears · claims check (Portia if in doubt) · no mid-word cuts, listened not assumed · captions in safe zones · right preset was on.
Ship it to Dougie. Final file + one paragraph: the angle, why, and what you changed by hand. That paragraph is your graduation.

7.2 How your work is measured

What	The bar
Quality first	Zero label errors, zero banned claims, zero harsh cuts reaching review. These outrank all volume numbers, always
Approval rate	Most of what you submit gets approved without rework, rising as your presets and teaching improve
Volume	Ramps with skill: dozens of finished pieces per week once fluent, because the agents carry the grunt work
Turnaround	Brief to first review same day; finished next day, most pieces
Teaching	At least a couple of committed improvements to agents, presets, or context every week

Full detail lives in KPIs-and-SOPs.pdf from your onboarding email. Targets calibrate in your first month; the quality line never moves.

7.3 What separates good from the best

the best editors here

Source their own briefs. They run Delphi and Ann without being asked and arrive with angles, not questions.
Verify like a skeptic, ship like an owner. Nothing "should be fine." They watched it, and they'd put their name on it.
Teach relentlessly. Their corrections land in files, so their taste slowly becomes the system's default.
Protect the brands. They know the never-say lists cold and treat compliance as craft, not bureaucracy.
Ask fast. Five minutes stuck means a message to Dougie or Dale. Ego waits; the queue doesn't.

7.4 Your operating rhythm, from here

When	What
Every session	git pull the brain, load your context, check what brands need
Daily	Workflow A on the queue; research ahead with B in parallel; finals to the Drive with everything approved
Friday	The 20-minute teaching loop from Day 6
Always	When unsure: don't ship, ask. Dougie for creative, Dale for technical

7.5 Final checklist

I completed the graduation run, all eight steps I sent Dougie the finished piece with my one-paragraph writeup I ran the full quality gate and can recite it I know how I'm measured and what the unmovable quality line is I know the five habits that separate the best, and which one I'll work on first

All done?