Loom Series — every recurring job, recorded once

Eleven Loom recordings Akash makes for Collins, one per recurring responsibility, each done live through Claude. For each: the context from the onboarding call, what to have open, the exact sentences to type into Claude, what to say, the skill that fires, and what done looks like.

Updated 2026-09-07

1 How this works

Eleven recordings, each one a recurring job, each done the way you actually do it: you talk to Claude, Claude does the work, you narrate why. Collins's job afterwards is to do the same thing by typing the same sentences. So every recording is built the same way:

Context. What he needs to know before watching, in your words from the call.
Have open. Tabs in order, so you never hunt for a page on camera.
Segments. For each: what is on screen, the sentence you type into Claude in bold, what you say in the quote box, and the skill that fires.
Done when. The artifact he will produce when he repeats it.

Rules on camera: name the page before you click it. Read aloud what you typed. When Claude answers, point at the one line that matters. No word without a one-line meaning; the glossary is at the end. Each recording under 40 minutes; split it if it runs long.

Record in this order. 0 and 1 tomorrow. 2 to 5 this week. 6 to 10 next week. He starts each project the morning after its recording.

#	Recording	Responsibility	Length
0	The business and your job	Context	15 min
1	The daily loop on AlphaLift	Delivery, every morning	40 min
2	The AlphaLift and Level King post-mortem	Delivery, why leads died	30 min
3	The conversation practice sheet	Conversations, the script	20 min
4	Reconcile our board against the client's CRM	Transfers and post-transfer	20 min
5	Dress every Facebook page	Pages	20 min
6	A page from created to launch-ready	Pages	30 min
7	What the ads did and how we made them	Content, reading performance	25 min
8	Build one ad: voice, picture, words	Content, the weekly pack	30 min
9	Post to every page	Content, organic	25 min
10	Comment to message, and the problems board	Conversations, issue tracker	25 min

2 Loom 0 · The business and your job (15 minutes)

Context. This is the call, condensed, so he can rewatch it in fifteen minutes instead of two hours.

Have open: the Start Here page, the cheat sheet, Business Manager's Pages list.

On screen	What you say
Start Here page	"We sell appointments to home-improvement companies. Foundation repair, crawl space, waterproofing, the high-ticket jobs, because one job is ten to fifteen thousand dollars to them. We find homeowners who need the work. I sell the appointment to a contractor in that area. That is the whole business."
Business Manager, Pages list	"Look at the page names. Not the company's brand. State plus the work: 'TX Foundation Repair & Waterproofing.' The page name is prime real estate. When someone scrolls, they see the video, then the page name, then the caption. 'Dallas Foundation Repair' tells them it is local and what it does in half a second. A brand name tells them nothing. One page per state, every market in that state inside it."
Same	"The old way: a client paid a retainer, then I built the page. The new way: I build every page now, post content now, run a little of my own money, and sell the leads. The setup happens before the client, not after. When a homeowner in Kansas gives me a phone number, I call every foundation company in Kansas and sell that inspection. Eight prepaid appointments for a thousand dollars, or three to four for five hundred. No retainer, no risk to them."
Cheat sheet, the six-area table	"Your job is everything except the selling. Six things. Pages: create them, dress them, keep them healthy. Content: the same video on every page, caption tweaked north versus south, one post a day. Conversations: every comment gets a message, every message gets a reply, the bot asks two questions max and then the phone number. Transfers: the moment we have a name and a phone, it goes to the client's CRM, formatted right. Post-transfer: did they call, quote, win. And the problems list: anything that breaks twice gets fixed for good."
Cheat sheet, the glossary	"None of this is manual. AI agents create the pages, post the content, answer the comments, send the leads. You oversee the agents. When something breaks, and things break constantly, you find it and fix the system, not the one lead. The way you do all of it is by talking to Claude. Everything I do, I do by asking Claude. You will too."
Same	"Two numbers to remember. Seventy-five percent of our leads should turn into a booked visit; the industry gets twenty-five. The difference is speed and a real conversation. And five minutes: that is how long you have to respond before a homeowner moves on."

Done when: he can say the six jobs and the two numbers back.

3 Loom 1 · The daily loop on AlphaLift (40 minutes)

Context. His every-morning job. Five screens, in order, done by 9:15. Recorded on a real morning with real leads.

Have open, in this tab order: Command Home, the nightly checks, the list view, the lead board, the client lane, Austin's DripJobs board (app.dripjobs.com), the problems board, Claude in the workspace, fresh session.

"This is my morning, every morning, and from next week it is yours. Five screens, in this order, done by 9:15. Each one is a question."

Segment 1 · Command Home. "Is anything red?" Point at one weekly number and one rail in the band. Red becomes a card on the problems board today.

Segment 2 · The nightly checks. Sort by nights running.

Type into Claude	Explain the top three findings on the nightly checks by streak, and which are about AlphaLift.

"Every night at 5:30 the database asks itself 35 questions. A finding two nights in a row is the signal. One night is noise."

Segment 3 · The list view. Did yesterday's leads land?

Type into Claude	For every lead since yesterday with a name and a verified phone, show me the delivery rows: client, channel, status, last error, sent time.

"This table is what our system believes. Now the truth." Switch to Austin's board, global search one name from the table. "Name, phone in the phone field, source St. Pierre. Proven. 'Sent' on our side is a timestamp. A card on his board is proof."

Pick a stuck one.

Type into Claude	Why was lead [name] not delivered? Read the send-guard log first, then the GoHighLevel execution log, then the contact. One sentence on the cause.

"That order, never skipped, never guessed. For a month we guessed and the guesses were wrong."

Type into Claude	Save yesterday's leads as three columns, proven, unproven, stuck, to the ops log for today.

Segment 4 · The lead board. Who is waiting on us? Open one "Needs reply" card, read the last three messages.

"What should the bot have said? Almost always: answer the question, then ask for the phone. Two questions, then the phone. Every extra question loses people."

Type into Claude	Show me the last ten send-guard log rows for lead [id]: when, kind, allowed, reason.
Type into Claude	Add a note to that card: cause in one sentence, next action, due tomorrow.

Segment 5 · The client lane. What is the client sitting on? Green under a day, amber under two, red past two. Open the AlphaLift coaching draft. Check Austin's board and calendar first.

Type into Claude	Rewrite the AlphaLift coaching draft as one message from me to Austin. Plain, short, end with: mark it in your portal, or reply with one word.

"Board first, ask second. Portal first, always: a portal tap updates our records by itself; a text needs a human to type it in."

Type into Claude	Log one client send in the daily touch log: AlphaLift, now, nudge on [name].

Segment 6 · The problems board. Filter to unowned.

Type into Claude	Log a clunk: key [short-key], what broke: [one line from this morning].
Type into Claude	Set owner to me on the card for the finding from segment 2, and read it back.

"The system finds it, a person owns it, the ledger records the fix. If you cannot say how you would undo a fix, it is not designed."

Done when: he runs all five screens alone and posts the three-column list and one client message by 3 pm.

4 Loom 2 · The AlphaLift and Level King post-mortem (30 minutes)

Context. Between the two clients: 263 conversations, 114 cards on client boards, 16 booked visits, 7 jobs won, all seven Level King's. The page you walk through shows every real lead and where it stopped. This is the recording that teaches him what "looking for problems" means.

Have open: the post-mortem, Austin's DripJobs board, Dan's DripJobs board, Claude.

On screen	What you say
The funnel table	"Read it top to bottom. Level King leaks between qualified and card: a month where 'sent' meant a text to my own phone. AlphaLift leaks between card and estimate: 39 of 43 cards never became a call, because the cards had no phone number and nobody nudged Austin. Same money, two different holes."
The timeline	"Every row here is a week we did not know something was broken. Read the last column. Each rule exists because we lost leads without it."
The ten holes	"These ten will happen again. Each has a check. Holes one, two, four and seven are daily. That is twenty minutes."
AlphaLift lead table	Pick three rows: one with a card and no estimate, one stuck with no card, one disqualified. For each, say what you see and what the fix would have been.

For the "card, no estimate" row:

Type into Claude	Show me everything we know about lead [name]: the conversation, the delivery rows, the card, and any outcome. Then tell me why it never became an estimate request.

"Claude reads the brain, the guard log and the board and answers in one paragraph. That question, asked for every red card, is the post-mortem you run every week from now on."

For the disqualified row:

Type into Claude	Why was [name] disqualified, and was that the homeowner's own words or our rule?

"Only four things really disqualify: STOP, wrong number, 'lose my number', and not English. Anything else, we probably closed too early. Kevin Ash was killed by a city list. He was a real lead."

| Level King table | Pick one won job and trace it backwards: card, estimate, booked, won. "This is what right looks like. Twenty minutes from ad to booked. Everything we do is trying to make that the normal case." |

Done when: Monday's project. One row per AlphaLift conversation with cause and recoverable yes/no, and an issue row for each recoverable lead.

5 Loom 3 · The conversation practice sheet (20 minutes)

Context. The bots drifted off script about half the time. On 2 September I had Claude pull every conversation pattern we handled badly, 25 of them, with what we sent and what we should have sent. I answer each one the way I would say it, and those answers become the script. This is how the script gets written: by me answering real messages, not by writing a document.

Have open: the practice sheet, Claude.

On screen	What you say
The top of the sheet	"Every card is a real homeowner message, what the bot said, how late it said it, and a default I would use. The rules at the top are the ones every answer must pass: one message per turn, never re-ask something they already told us, answer their question before asking yours, STOP is final."
First five cards	Type your answer in each box, out loud. Say why. "'Are you a bot?' I say: I'm the page admin, a real crew does the estimate. One line, then back to the phone number."
"Copy all answers"	"This is the loop. I answer, Claude turns the answers into templates, the bot uses the templates. Until my answers are in, the bot improvises."

Type into Claude	Here are my answers to the first five practice-sheet patterns. Turn each into a message template, check it against the copy rules, and tell me which ones already exist.

"Copy rules: no adjectives, no adverbs, first person, 'sales rep' never 'estimator', never 'I run ads'. Claude checks every line against those before it becomes a template."

Type into Claude	Audit the last 20 AlphaLift conversations against the script and show me every reply that went off script, with the line it should have been.

"That audit is weekly. The skill that runs it reads the live bot and compares it to the approved copy."

Skills that fire: ghl-conversation-ai.

Done when: his own five answers pasted to you, and one audit saved to the ops log.

6 Loom 4 · Reconcile our board against the client's CRM (20 minutes)

Context. From the call: "we sent over 50 leads and realized they weren't even sending, then that they went without a phone number." The only way to catch that is to look at the client's side every day. This recording is that check, on both clients.

Have open: the list view, Austin's board, Dan's board, Claude.

Type into Claude	Compare our transferred AlphaLift leads from the last 14 days with the cards on Austin's board. List any lead we say is sent that has no card, and any card on his board that we cannot match to a lead.

"Two lists. The first is leads we lost on the way. The second is cards we cannot see the outcome of, because the card came back without our lead id. Both are problems; the first is urgent."

On screen	Austin's board, pipeline view

"Estimate Requested and everything to the right is what we want. Everything to the left is a card sitting there. Anything from us is marked source St. Pierre. Google ones are his own ads, not ours."

Type into Claude	For every AlphaLift card older than 48 hours with no estimate request, draft one line per lead for the message to Austin.

Repeat once on Dan's board.

Type into Claude	Same check for Level King: our transferred leads against Dan's board, last 14 days.

"The client's CRM is the truth for outcomes. Our labels lie both ways: we had a lead marked 'declined by Dan' that had an estimate booked, and one marked 'booked' that was not on his calendar. Board before ask, every time."

Done when: the two lists saved to the ops log, one line per stuck lead, and an issue row for every unmatched card.

7 Loom 5 · Dress every Facebook page (20 minutes)

Context. From the call: "the ones with profile pictures versus the ones without: all of that is automated. I ask Claude which pages have pictures and banners and which don't, and for the ones that don't, add them." Six or seven pages were created in the last few days and have nothing on them.

Have open: Business Manager's Pages list, Claude, one dressed page in Chrome (Texas) as the reference.

On screen	What you say
Pages list	"Look at the row of pages. The ones with a picture were dressed by Claude. The grey silhouettes were created and never dressed. A page with no picture runs ads as a blank; people do not message a blank."
The Texas page	"This is the standard. Cover rendered from a template, bio with the state name, first post. Every page looks like this with the state swapped."

Type into Claude	Which of our state pages are missing a profile picture, a cover or a bio? List them.
Type into Claude	Dress every page on that list with the foundation repair cover and bio for its state.

"Claude renders each cover from the template, then pushes cover, bio and badge to the page through Meta's API with our page token from the vault. No browser needed for this part. It takes about a minute a page."

On screen	Refresh one of the newly dressed pages in Chrome.

"Check one by eye every time. Rendering is deterministic; Meta's caching is not."

Type into Claude	Read back the live name, picture and cover status for every state page.

Skills and tools that fire: the branding scripts (render_covers.py, apply_branding.py), the vault's page tokens. No skill trigger phrase; Claude finds the scripts by the sentence.

Done when: every page in Business Manager has a picture, cover and bio, and the readback shows no gaps.

8 Loom 6 · A page from created to launch-ready (30 minutes)

Context. On the call we created the Kansas page together and stopped there. A page is not usable until six more things are true, and every one of them has failed silently at least once. This recording finishes Kansas.

Have open: Business Manager, Events Manager, GoHighLevel, Claude, the cheat sheet, Area 1.

Type into Claude	Run the page readiness sweep and show me everything Kansas is missing.

"Six things make a page real: our two service accounts attached, its own tracking dataset proven by a test event, the Lead Ads terms accepted, the page connected in GoHighLevel, and the four database rows that route its leads. Missing any one and leads are written correctly and delivered to nobody. No error. Seven markets ran ads that way for a week."

Then, in order, saying which are clicks and which are Claude:

1The rename to "KS Foundation Repair & Waterproofing". Your click; Meta refuses renames by API. "Sixty-day lock after, so get it right."
2Attach both service accounts to the Kansas page. Claude, through the API.
3The tracking dataset in Events Manager. A browser step. "Claude drives Chrome to the screen; the API route makes the wrong kind of object, we proved that in August."
4Fire the dataset test event for the Kansas page and show me the result. Claude. "'Received' is the only proof the page and the dataset are paired."
5Lead Ads terms. Your click at the terms link. "Per page. Not once for the business."
6Connect the page in GoHighLevel, both the lead-form sync and the Messenger connection. Your login. "Say this out loud in every launch: is the page connected? Seven markets, zero leads, because it was not."
7Stand up Kansas's routing rows, dry run first, then apply. Claude. "It refuses without the test-event proof from step 4."
8Add the Kansas page to the router and deploy through the gate. Claude.
9Run the page readiness sweep again.

"Green on every line. No ads yet; ads wait for a buyer. But the day a Kansas contractor says yes, this page is live in an hour."

Skills and tools that fire: the readiness sweep, the stand-up script, the router deploy through the gate, Claude in Chrome for the two browser steps. Later: new-launch-qa ("run the launch QA on Kansas").

Done when: Tuesday's project. The readiness sweep shows Kansas green on every check.

9 Loom 7 · What the ads did and how we made them (25 minutes)

Context. He has never run ads and does not need to. He needs to read them: which video won, why, and what changes next. Everything we make is a variation of a winner.

Have open: the scoreboard, Creatives tab on All time, the Markets tab, Drive "Proven B2C Ads by Niche", Claude.

On screen	What you say
Creatives tab, sorted by cost per qualified	"One row per video. The only number I sort by is cost per qualified lead. Views and clicks are noise. Two things to note on a winner: the first words of the voice, and what is on screen in the first second. The hook and the surface."
Play the top video	"Forty-one words. Who we are, what we lift with the surface named first, without the cost or mess of replacement, send us a message. No adjectives. Red captions, white words, the spoken word turns yellow. This format beat silent loops two to one on cost per conversation."
Markets tab	"Same money by city and client. Spend, leads, cost per lead, replied, qualified, sent to client, appointments, won. A market spending with zero replies usually means the page is not connected."

Type into Claude	Rank the top five foundation repair videos by cost per qualified lead, all time. For each: hook words, surface on screen, spend, qualified, cost per qualified, and which markets it ran in.
Type into Claude	What do the top three have in common, and what is the one thing to change for the next pack?

"That answer is the brief. The pack is the bet. Never fewer creatives: if per-ad budget looks thin, raise budget or extend the window, never cut the number of videos."

| Drive, proven ads folder | "Finished winners live here at full size. Before we make anything new, we re-voice one of these. New footage costs money and time; a proven winner with a new read is the cheapest test we have." |

Done when: he can name the top three hooks and the one change for next week without looking.

10 Loom 8 · Build one ad: voice, picture, words (30 minutes)

Context. Once a week a new set of foundation repair videos gets built and staged for my approval, even while ads are off, because the day a client signs we relight the same day. Three parts to every ad, all made by Claude.

Have open: Voicebox running, Claude with a fresh session named "FR pack week 37", Drive proven ads.

Type into Claude	Write the brief for this week's FR pack from the top three hooks, which proven ad to re-voice first, and the market. Save it in ready-to-deploy as this week's brief.

The voice.

Type into Claude	Generate an Eric voiceover for this script with the standard treatment: We are a local company that specializes in foundation stabilization. We fix sagging floors, wall cracks and sticking doors, without the cost or mess of a rebuild. Send us a message. We can come out and get you a free estimate right away.

"Eric is our voice, a free local voice engine on this computer. The treatment is a slight speed-up and the pauses cut short, with one longer pause before 'send us a message'. Raw computer voice never ships. The gate compares the pace to the read that converted; fail means regenerate. We never stretch the voice to fit a clip. The picture gets retimed to the voice."

The picture.

Type into Claude	Generate one five-second foundation-crack clip with Higgsfield: vertical, no text, no people, a real crack in a concrete slab.

"It shows the credit cost before it renders. Always look at that number. Real footage beats generated on cost per lead; generated is for gaps."

The words and the sheet.

Type into Claude	Burn the red karaoke captions from the script onto that clip, loop the clip to the length of the read, and build a contact sheet for this one ad.

"Captions come from the script, never from listening to the audio: the listener hears 'live' for 'lift'. The contact sheet is the approval. Nothing goes to Meta before I say yes to the sheet."

The steps he watches but does not do this week: create the ads paused, run the launch check, activate on my word.

Skills that fire: b2c-creative-production ("build the FR pack", "add the voiceover", "burn the captions"), higgsfield-generate ("generate a clip"), broll-tos-pipeline ("clean these clips", "stage crawl space for tos").

Done when: Thursday's project. A brief, a treated read, one clip, one captioned ad, one contact sheet in your hands.

11 Loom 9 · Post to every page (25 minutes)

Context. From the call: "the content on every page can be exactly the same; only the page name changes, and the caption tweaks north versus south." Nothing organic has ever been posted to a homeowner page. This recording does the first one, and the notes from it become the specification for the posting system.

Have open: Business Manager Pages, one live state page in Chrome, Drive proven ads, Claude.

Type into Claude	List the live state pages and which of them our page tokens can post to.

"Posting goes through the page token, the same way dressing does. If a page is not in that list, the token cannot see it and the fix is in Business Manager, not here."

Type into Claude	Take the top proven foundation repair video. Write two captions, one for northern states about freeze and thaw, one for southern states about soil and heat. End both with: comment 'quote' for a free inspection. Post it to every live page with the right caption. Log page, caption and post id.

"The call to action on every organic post is the same: comment to get a quote. That is what Loom 10 turns into a message. Same video, forty pages, two captions. Volume first. We learn what works from what gets comments."

On screen	Refresh one page in Chrome; show the post.

Type into Claude	Write a note titled "what broke" with everything that failed or needed a workaround in that posting run, and save it beside the log.

"That note is the spec. The posting system does not exist yet. It gets built from what breaks the first time."

Done when: Wednesday's project. One post per live page, the log, the "what broke" note.

12 Loom 10 · Comment to message, and the problems board (25 minutes)

Context. The comment automation already exists, built in July, on for one client only. It reads a comment, sends a private message, then posts "check your DMs". After the DM, the bot takes over and asks for the phone. This recording turns it on for one page, tests it once, and ends on the problems board where every issue from the week lands.

Have open: the Kansas page in Chrome, a test account logged in on a second browser, the lead board, the problems board, Claude.

Type into Claude	Show me how the comment-to-message automation is set up for Level King and what the Kansas page needs to run it.

"Three things per page: the page token in the worker, the page subscribed to comments, and the client's flag turned on. And one rule to fix: today it refuses to message anyone who posts a phone number in the comment. Under the new plan that person is the lead."

Type into Claude	Turn comment-to-message on for the Kansas page only, with the standard script.

From the test account, comment "quote" on the Kansas post.

Type into Claude	Show me the message that went out to that commenter and the board card it created.

"Comment, message, phone, card. That is the entire front of the funnel. Watch the board card appear. Then turn it off again until I rule on fleet-wide."

Type into Claude	Turn comment-to-message off for Kansas.

The problems board.

Type into Claude	Show me every unowned card on the problems board.
Type into Claude	Set owner to me on these ten. Log a clunk for each thing that broke this week. Rebuild the board and deploy it through the gate.

"Every problem once. Three times in a week is a build, not a fourth patch. If you cannot write how to undo a fix, it is not designed."

Skills that fire: the comment automation in the sales-crm worker, publish-page for the board redeploy through the gate.

Done when: Thursday and Friday's projects. One test comment produced one message and one card; zero unowned cards; five or more clunks; one ledger row per fix.

13 The words, one line each

Word	Means
The brain	Our database. Every lead, message and change lives there. Claude reads it.
The vault	The part of the brain that holds passwords and API keys. Claude uses them; you never copy one.
The ledger	The log of every change: what, why, how to undo it.
The send guard	The gatekeeper on every automated text. Says yes or no and writes down why.
The nightly checks	35 questions the database asks itself every night. A repeat is the signal.
The problems board	One card per thing that keeps breaking.
A clunk	One line about something that broke today. Three in a week becomes a build.
DripJobs	The clients' own job boards. The only proof a lead reached them.
GoHighLevel	Our messaging system. The bot and the texts live there.
The page token	The key that lets Claude act on a Facebook page without a browser.
The contact sheet	One page of every ad in a pack. Akash approves it before anything ships.
A skill	A packaged how-to Claude follows when it hears the trigger words.