I stole my competitor's $1m/d framework and I want you to do the same.

You have control of my actual Chrome browser through the Chrome connector.
I am already logged into Meta Ads Library, Google Ads, Shopify and Google Drive.

Standing rules for this session:
- Use my browser directly. Do not ask me to paste screenshots, export CSVs
  or hand you data you can go get yourself.
- If a page needs scrolling to load more results, scroll it. If it needs a
  click, click it. Do not stop and ask me to do it.
- If something fails, tell me exactly what failed and what you tried. Do not
  silently work around it with a worse method.
- Save every artifact you produce to disk in the working folder so we can
  rebuild without re-scraping.

Confirm you can see my browser and list the tabs that are currently open.

Go to the Meta Ad Library in my browser and pull the active video ads for
[COMPETITOR PAGE NAME].

1. Sort by impressions, highest first. Scroll until you stop getting new
   results, at least 15-20 scrolls.
2. For every video ad, capture: the ad ID, the page, the product line,
   days running, the number of ads sharing that creative (variations),
   and the video source URL.
3. Download every unique video to ./library/. Pull the file server side
   from video.currentSrc, the in-page URLs are signed and masked.
4. Dedupe by md5. One creative = one hash. Sum duplicates into a
   "variations" count. Tell me how many unique creatives came out of how
   many listings.
5. For every unique creative, measure with ffmpeg and ffprobe:
   - runtime in seconds
   - hard cut count using: select='gt(scene,0.3)'
   - cuts per minute, average shot length, cuts in first 10 seconds
   Use scene threshold 0.3 and nothing else. We reuse it later to score
   our own ads and the comparison has to be apples to apples.
6. Transcribe the top 30 by days running with whisper.cpp. Only the top 30.
   Record word count and words per minute for each.
7. Build a 12 frame contact sheet per ad (tile=4x3) and actually look at
   each one before you tag it.
8. Tag every ad: format, character archetype, set, hook type, hook visual,
   mechanism, offer device, caption style, edit style.

Output a single JSON file with one record per unique creative.
Flag anything you could not measure rather than guessing.

Using the teardown JSON, build me a Google Spreadsheet called
"[BRAND] CREATIVE TRACKER" with these tabs. Use the Drive connector to
create it, then Apps Script to format it, because the connector can create
files but cannot edit cells.

TAB 1 "BENCHMARKS"
One row per unique competitor creative, sorted by days running descending.
Columns: days running | runtime (s) | runtime bucket | variations |
library copies | page | product line | format | hook (first words) |
words | wpm | Ad Library link | status
Status vocabulary: WINNER 60+ / WORKING 30+ / TEST under 30.
Buckets: under 30s, 30-90s, 90s-4min, 4-8min, 8min+.

TAB 2 "DASHBOARD"
A runtime KPI table: for each bucket give creatives, % of creatives,
% of scaled variations, % running 30+ days, % running 60+ days, longest
running, and count of 60+ day winners.
Then a table "formats of the 60+ day winners" counting each format.
Then four charts: share of creatives by length, % still running at 30/60
days by bucket, where the 60+ winners come from, formats of the winners.
Put a 4 tile KPI strip at the top with the winning bucket, its share, the
longest running ad, and the survival gap between the best and worst bucket.

TAB 3 "VISUAL + CONCEPTS"
Same creatives, different columns, read from the contact sheets:
library ID | link | page | product line | days running | variations |
runtime | hard cuts | avg shot (s) | cuts per min | edit style | concept |
avatar + set | hook visual | caption style | opening line
Edit style buckets: static anchor (3 cuts or fewer), light cutting
(12 or fewer per min), montage (over 12 per min).
Then a KEY FINDINGS block at the top, numbered, each one a sentence I could
act on, each one backed by a number from this tab.
Then an EDIT STYLE vs LONGEVITY table, and a per product line table
comparing their mature line against the line that competes with me.

TAB 4 "CREATIVES"
My production queue. Colored bands by angle type in this order:
ROOT CAUSE, STRONG CALL OUT, TOPIC, COMPARISON, OFFER, SOCIAL PROOF,
MYTH BUST. Repeat the header row under each band.
Columns: Ad ID | Angle Type | Angle/Title | Hook (first 3 seconds) |
Format | Target runtime (KPI) | Script + Copy (Doc link) |
Video File (Drive link) | Status | Destination | Notes
Status dropdown: SCRIPTED, IN PRODUCTION, READY, TESTING, WINNER, KILLED.
Target runtime values must come from the DASHBOARD buckets, never invented.

Make a hidden BACKUP tab before any future reorganisation.

Write a [RUNTIME] direct response video script for [PRODUCT], format
[FORMAT from the dashboard], angle [ANGLE].

Pull the format, runtime and word count target from the DASHBOARD tab.
Do not invent a runtime.

Brief it first, in this order, and show me this before the script:
avatar / enemy list / hidden cause / mechanism line / messenger /
proof timeline / vindication / objection / offer / closing line.

Then write the narration against this beat map, as % of runtime:
  0-3%    Hook. Twist, repetition, or a shocking specific.
  3-15%   Symptom mirror in the avatar's own words
  15-30%  Wound + enemy list + "it is not your fault, it is not your age"
  30-45%  Messenger arrives, hidden cause revealed
  45-55%  Mechanism, and the product physically enters for the first time
  55-60%  Objection, then permission
  60-85%  Proof timeline, a felt change at every step
  85-92%  Vindication, callback to the hook
  92-100% Offer, guarantee, CTA, closing line

The hook must contain all four parts:
  1. public humiliation by someone intimate, witnessed by others
  2. one absurdly specific detail (a number, an age, a duration)
  3. the withheld save from a mismatched source (not a doctor)
  4. the rewind line that buys the runtime

The product name appears NOWHERE before the 45% mark.

Writing rules:
- 4 to 10 word sentences. Fragments are fine. Contractions always.
- One concrete noun or number per sentence where you can.
- Repeat the hook line and the mechanism line 3x across the ad.
- Blame external enemies, never the viewer.
- Name the viewer's private behaviours (oversized shirts, no photos,
  on the couch by seven).
- Put the joke immediately before the objection or the ask.
- Hard ingredient names get written on the label, never spoken.
- No em dashes. No AI words: journey, unlock, game changer, elevate,
  delve, landscape, testament.

Deliver:
1. The narration, with % markers in the margin
2. A shot list, one shot per sentence
3. The Meta primary text, written CALM. All the drama stays in the video.
   Headline under 40 characters.
4. A variation plan: 5 alternate hooks and 3 alternate characters or
   settings for this same body script. Write these now, not later.

Split this script into generation takes. Rules:

- Split at sentence boundaries first, then commas if a sentence is too long.
- Maximum 24 words per take. Never break mid clause.
- Assign duration by word count:
    4 seconds  if 11 words or fewer
    6 seconds  if 12 to 18 words
    8 seconds  if 19 to 24 words
- 8 seconds is a hard cap. If a sentence will not fit, split it.

Then print:
- the take list with word count and assigned duration
- total runtime
- estimated words per minute (target 170)
- estimated credit cost, using 4s=11, 6s=16.5, 8s=22, and DOUBLE it
  because we generate two variants of every take

Show me the cost before you generate anything.

Step 1, the still. Generate the character with Nano Banana Pro at 2k,
in the aspect ratio of the final ad. Cheap, so reroll as many times as
you want. Give me 4 options.

Character: [AGE, REGION, BUILD, WARDROBE, SETTING, PROP, EXPRESSION]
Style: photoreal, natural available light, flat ungraded phone camera
color, no cinematic grade, no ring light. She/he does not look at camera.

Once I approve one, save its job ID as THE character reference and use it
as start_image on every single anchor take from now on. Never regenerate
the character from text again, in this ad or any future variation.

Step 2, the voice block. Write ONE block covering:
age, region, timbre, volume, vowel shape, inflection direction, attitude,
and pace in words per minute (use 170).
Then append this delivery note verbatim:
"every word is pronounced as one ordinary spoken word, do not spell any
word out, do not break any word into separate syllables, do not
over-enunciate."

Save the voice block to voice.txt. Paste it byte identical into every
take prompt. Do not rewrite, tighten or improve it between takes.
Before generating, md5 the voice block in every prompt and confirm they
all match.

Generate this take on Higgsfield with:
  model veo3_1, variant veo-3-1-fast, quality basic
  duration [4 / 6 / 8], aspect [9:16 or 16:9]
  start_image = [the locked character still, or previous take's last frame]
  count 2   <-- always two variants
  declined_preset_id = [the preset id]  <-- or the tool returns a preset
                                            recommendation instead of a job

PROMPT BODY, in this order:

No subtitles. No captions. No music.

SHOT - [SHOT DESCRIPTION]. Static camera, identical framing to the
provided image for the entire clip, no camera movement. Generic unbranded
props, no brand names or logos on any objects.

VOICE - IDENTICAL IN EVERY SHOT, DO NOT VARY: [paste voice.txt verbatim]

Says, word for word, and must say the complete sentence all the way
through to the final words: "[LINE]"

SOUND: voice only, recorded in a quiet room. Absolutely no music, no
score, no instrumentation of any kind at any point in the clip. No sound
effects, no animals, no clock, no birds, no wind. Nothing but the voice.

Write a [GENRE] story song for a [RUNTIME] ad. The whole ad is sung,
there is no spoken narration. The verses carry the plot, the chorus is
the sales argument.

Structure it to the same beat map as a spoken script. Product name does
not appear until the 50% mark, which lands at the bridge.

Pacing rule that matters more than anything else:
give every line about 5.5 seconds. Four lines per verse in a 22 second
section. Cram 8 lines into 20 seconds and the vocal turns to mush.

Protect any name in the chorus with a hard consonant in front of it.
"about Earl" renders as "about her". "his name is Earl" survives.

Deliver three blocks I can paste straight into Suno:
1. STYLE OF MUSIC box: genre, bpm, vocal description, instrumentation,
   and the words "clear enunciated storytelling, vocal front and centre"
2. EXCLUDE STYLES box: mumbled vocals, slurred words, heavy reverb,
   vocal buried in the mix
3. LYRICS with [Verse] [Chorus] [Bridge] tags

Then: generate the chorus on its own first, 30 seconds. If the hook does
not stick, nothing after it matters and you have wasted nothing.

After I bring back a track, transcribe it with whisper before we build
anything against it. If the words are not legible in the transcript they
are not legible in the ad.

From now on, for any multi take ad, chain the frames. Never generate all
takes from one shared still.

Procedure:
1. Take 1 uses the approved character still as start_image.
2. After each take is approved, extract its LAST frame:
     ffmpeg -y -v error -sseof -0.25 -i take.mp4 -update 1 -q:v 2 last.jpg
   -sseof seeks from the END of the file. -update 1 keeps overwriting so
   the final frame written is the one left on disk.
3. That frame becomes start_image for the next take.
4. Before chaining from any frame, check it is not a fade to black:
     mean luma < 12 of 255 = do not chain, regenerate that take instead
5. Repeat to the end of the ad.

Constraints you must plan around:
- Chained takes CANNOT be batched. This is strictly serial. Tell me the
  expected wall clock before starting.
- If a take fails QA, everything after it has to be regenerated too,
  because the chain is broken. Approve each take before continuing.

Exception: if a take deliberately ends on a white out or a flash, that is
fine and actually useful. Check the luma, and if it is very high, tell me,
because we can hide the next cut inside it.

For every generated take, run both checks and give me one table.

1. Transcribe it:
   ffmpeg -i take.mp4 -ar 16000 -ac 1 t.wav
   whisper-cli -m ggml-base.en.bin -f t.wav -nt -np

2. Contact sheet it:
   ffmpeg -i take.mp4 -vf "fps=1/3,scale=180:-1,tile=5x2" -frames:v 1 sheet.jpg

Then report per take:
- match ratio of the transcript against the scripted line (difflib)
- measured words per minute of the actual speech
- the timestamp the speech ENDS vs the clip duration
- whether the character looks on-model in the contact sheet

FAIL any take where:
- a word from the script is missing
- speech ends within 0.15s of the clip end (it will clip in the render)
- measured wpm is under 140 (padded, it will feel slow)
- the character design drifted

For failures, tell me which of the two variants to use, or say both are
dead and it needs a reroll. Mishearing a proper noun is fine, captions
come from the script not from whisper.

Assemble the approved takes into the final master.

STEP 1, NORMALISE EVERYTHING FIRST. Takes come back with different sample
aspect ratios, frame rates and loudness. xfade refuses to join mismatched
streams and mismatched loudness is audible at every cut.
  -vf "scale=[W]:[H]:force_original_aspect_ratio=increase,crop=[W]:[H],
       fps=24,setsar=1,format=yuv420p"
  -af "afade=t=in:st=0:d=0.03,afade=t=out:st=[end-0.03]:d=0.03,
       aresample=48000"
Those 0.03s fades kill the clicks at the joins.

STEP 2, CUT ON WORDS NOT TIMECODES. For each take, align the whisper
transcript to the script with difflib, then trim 60ms before the first
word and 220ms after the last, snapped to 24fps frame boundaries.
Specify any b-roll as "cut to this clip on word N", never as a timestamp.
When the voice moves, the edit moves with it.

STEP 3, JOIN. Compute crossfade offsets from PROBED durations, never
assumed. 0.10s dissolve for frame chained takes, 0.18s for unchained.
acrossfade on audio with c1=tri:c2=tri.

STEP 4, CAPTIONS, burned in off word level whisper timings.
- 1 to 3 words per card, bold white, black outline 3-4px
- ONE keyword per card in yellow (#FFE600)
- flush the card at 3 words, or a gap over 0.35s, or sentence punctuation
- keep captions off faces and off the product
- captions come from the SCRIPT, not the transcript

STEP 5, SOUND. One music bed across the WHOLE ad, never per take. This is
the single thing that stops stitched clips sounding stitched.
  [voice]asplit=2[vox][key];
  [bed][key]sidechaincompress=threshold=0.04:ratio=9:attack=15:release=350[duck];
  [vox][duck]amix=inputs=2:normalize=0[mix];
  [mix]loudnorm=I=-14:TP=-1.5:LRA=11[out]
Bed sits 18-24 dB under the voice, ducked by the voice itself. Room tone
under everything at -45 to -55 dBFS.

STEP 6, EXPORT. H.264 High, 24fps, CRF 19, yuv420p, AAC 192k at 48kHz,
+faststart. Name it [BRAND]-[PRODUCT]-AD[nn]-[slug]-[aspect]-v[n].mp4

STEP 7, QA THE MASTER before you hand it to me:
- hook line and hook visual land in the first 2 seconds
- no black or blank first frame
- cut density matches the format (talking head = 3 hard cuts max)
- no narration gap over 0.4s, no pops at the seams
- one voice timbre the whole way through
- captions synced within 100ms
- product label correct and not mirrored
- offer card on screen at least 3 seconds
- measure integrated loudness and confirm it is within 1 LU of -14

Build me an AD SCORECARDS tab in the creative tracker, and score this ad.

MEASURED (35% of total), all automated:
- runtime, and which dashboard bucket it lands in
- cuts per minute, using the SAME scene threshold 0.3 we used on them
- words per minute from the whisper transcript
- motion energy: median signalstats YAVG on a tblend=difference at fps=4,
  scaled to 90x160, dropping hard cuts
- hook visual change: is there a cut or a major frame change before 5.5s
Score each against the band the 60+ day winners occupy in the BENCHMARKS
tab. Do not use fixed bands, derive them from the winners.

RUBRIC (65% of total), 0-10 each, judged from frames placed side by side
with the real competitor ads:
hook visual / avatar and set / phone real framing / performance energy /
captions and badge / script structure / offer clarity / AI realism

Then score 3 to 5 NAMED competitor ads on the exact same rubric, and
output a leaderboard with ours inserted into it. Include their days
running and their variation count next to each.

Columns: Ad | Title | Competitor format modeled | Runtime | Cuts |
Cuts/min | WPM | Energy | Measured /100 | Rubric /100 | TOTAL |
Rank vs competitors | Library percentile | Weakest scenes (reroll first)

The weakest scenes column is the actual output. Name the specific takes
and why: "take 7, 186 wpm, too fast" / "take 10, energy 5.76, slow".

If our ad loses, say it loses. A scorecard that always flatters us is
worthless.

Build me a "[BRAND] URL MAP" sheet with these bands:

ALL URLS
Every surface with: URL (production), preview/editor URL, purpose,
status, owner, notes. Statuses: LIVE, DRAFT, ARCHIVED.

AD-ONLY URLS (ad -> presell -> PDP -> checkout)
One row per angle. Columns: offer | angle/slug | presell URL (the link in
the ad) | PDP destination (where the presell CTA goes) | status | format |
notes. The 1:1:1 rule, one angle maps to one presell maps to one PDP.

ADS READY URLS
The whitelist. Only these four-ish rows may ever be pasted into Ads
Manager. Columns: destination | format/angle | website URL |
URL parameters | status | where it goes | best for.
Put the literal UTM string in the cell, ready to copy:
utm_source=facebook&utm_medium=paid_social&utm_campaign={{campaign.name}}
&utm_term={{adset.name}}&utm_content={{ad.name}}&utm_id={{campaign.id}}

STORE WIRING
Product IDs, variant IDs, selling plan IDs, one-time vs subscription
prices, first order vs refill price, and which template each uses.

ASSETS
Every image with where it is used and any per-image prohibition.

RULES
The hard rules, written as rules not suggestions.

OPEN ITEMS
Owner assigned, with status. Owner is me or you.

CHANGELOG
Date, surface, change. Append to this every time you touch anything.

ENFORCE THIS: when I ask you to script an ad, check the destination
exists in ADS READY URLS with status LIVE first. If it does not, stop and
tell me, do not script the ad.

File this finished ad:
1. Add a row to the CREATIVES tab under its angle band.
2. Create a Google Doc named "AD-[nn] · [title]" containing: a back link
   to the tracker, the info table (angle, format, runtime, hook, status,
   destination, UTM), the full script as a table (time | scene |
   voiceover + captions | cutaway), the AI generation prompts used, the
   Meta copy, and a pre-launch checklist.
3. Create a Drive folder for the video files and link it in the row.
4. Log runtime and credits spent. Set status READY.

NOW THE VARIATIONS. A variation is NOT a new ad.

An ad is a folder of takes where take 1 is the hook. To make a hook
variant:
  - hardlink every take except T01 from the base ad (os.link, zero bytes,
    not a copy)
  - generate ONE new T01
  - swap takes[0]['text'] in the config
  - leave takes 2 onward byte identical
Only the first 8 seconds are ever regenerated.

For the 16:9 version, do NOT re-render. Composite the existing vertical
take onto an outpainted plate with a feathered alpha mask, then rewrite
the config canvas and caption position.

At 30 days profitable, mark the ad WINNER and spawn 3 to 5 variation rows
using the alternate hooks already written in the script's variation plan.

Report how many delivered files came out of how many generated takes.

Set this up as a scheduled task, daily.

Every morning:
1. Re-pull the active library for each competitor in my watch list.
2. Diff against yesterday's pull and report ONLY the changes:
   - new creatives that appeared
   - creatives that died, and how many days they made it
   - creatives that crossed 30 or 60 or 90 days since yesterday
   - any creative whose variation count jumped, that is them scaling it
3. Re-measure anything new: runtime, cuts, cuts per minute, wpm.
4. Append new rows to BENCHMARKS, update the DASHBOARD tables, and if a
   runtime bucket or edit style changed rank, say so loudly at the top.
5. Update the CHANGELOG.

Then give me a five line summary, nothing longer:
- what they launched
- what they killed
- what they are scaling right now
- whether our house format defaults should change
- anything that beat our current best scorecard total

Do not re-transcribe ads we already have. Only transcribe new entrants in
the top 30 by days running.

function formatTracker() {
  var ss = SpreadsheetApp.openById('YOUR_SHEET_ID');
  var name = 'CREATIVES';

  // delete and rebuild so the script is idempotent. run it as often as you like.
  var old = ss.getSheetByName(name); if (old) ss.deleteSheet(old);
  var sh = ss.insertSheet(name, ss.getSheets().length);

  var G = '#0c3b2f',   // band green
      H = '#e8ece9',   // header grey
      W = '#ffffff';

  // this one line is why it stops looking like a spreadsheet
  sh.setHiddenGridlines(true);
  sh.getRange('A:K').setFontFamily('Inter').setFontSize(10)
    .setVerticalAlignment('middle');

  var r = 1;

  // a full width dark band, used to open every section
  function band(text) {
    sh.getRange(r,1,1,11).merge().setValue(text)
      .setBackground(G).setFontColor(W).setFontWeight('bold').setFontSize(11);
    sh.setRowHeight(r, 26); r++;
  }

  // the light header row that repeats under every band
  function header(vals) {
    sh.getRange(r,1,1,vals.length).setValues([vals])
      .setBackground(H).setFontWeight('bold').setWrap(true);
    r++;
  }

  // title
  sh.getRange(r,1,1,11).merge()
    .setValue('[BRAND] CREATIVE TRACKER  ·  every ad, by angle')
    .setBackground(G).setFontColor(W).setFontWeight('bold').setFontSize(14);
  sh.setRowHeight(r,34); r += 2;

  var COLS = ['Ad ID','Angle Type','Angle / Title','Hook (first 3 seconds)',
              'Format','Target runtime (KPI)','Script + Copy','Video File',
              'Status','Destination','Notes'];

  // one band per angle type, header repeated under each
  var BANDS = [
    ['ROOT CAUSE: explains WHY people feel this way (the hidden reason)', ROOT_ROWS],
    ['STRONG CALL OUT: pattern-interrupt hooks that name a frustration', CALLOUT_ROWS],
    ['TOPIC: educational angle on one mechanism', TOPIC_ROWS],
    ['COMPARISON: us vs the alternatives', COMPARISON_ROWS],
    ['OFFER: price, guarantee, urgency', OFFER_ROWS],
    ['SOCIAL PROOF: reviews, UGC, testimonials', PROOF_ROWS],
    ['MYTH BUST: knocks down a common belief', MYTH_ROWS]
  ];

  BANDS.forEach(function(b) {
    band(b[0]);
    header(COLS);
    if (b[1].length) {
      sh.getRange(r,1,b[1].length,COLS.length).setValues(b[1]);
      r += b[1].length;
    }
    r++;  // breathing room between bands
  });

  // status colors, case sensitive so WINNER never matches "winner-ish" text
  var statusCol = sh.getRange(2, 9, r, 1);
  var rules = [
    ['WINNER',        '#b7e1cd'],
    ['TESTING',       '#fce8b2'],
    ['READY',         '#c9daf8'],
    ['IN PRODUCTION', '#efefef'],
    ['KILLED',        '#f4c7c3']
  ].map(function(p) {
    return SpreadsheetApp.newConditionalFormatRule()
      .whenFormulaSatisfied('=EXACT($I2,"' + p[0] + '")')
      .setBackground(p[1]).setRanges([statusCol]).build();
  });
  sh.setConditionalFormatRules(rules);

  // a real dropdown, so nobody types "in prod" and breaks your filters
  statusCol.setDataValidation(
    SpreadsheetApp.newDataValidation().requireValueInList(
      ['SCRIPTED','IN PRODUCTION','READY','TESTING','WINNER','KILLED'], true
    ).build()
  );

  sh.setColumnWidth(1,70); sh.setColumnWidth(3,220); sh.setColumnWidth(4,320);
  sh.setFrozenRows(2);
}

// tblStart is the row where your header sits, then 3 data rows under it
var chart = sh.newChart()
  .setChartType(Charts.ChartType.COLUMN)
  .addRange(sh.getRange(tblStart, 1, 4, 1))    // labels
  .addRange(sh.getRange(tblStart, 4, 4, 2))    // the two % columns
  .setNumHeaders(1)
  .setOption('title', '% still running at 30 / 60 days, by edit style')
  .setOption('colors', ['#8fc9b3', '#0c3b2f'])
  .setOption('legend', { position: 'top' })
  .setOption('vAxis', { format: 'percent' })
  .setPosition(r, 1, 0, 0)
  .build();
sh.insertChart(chart);

// percentages have to be formatted as percentages or the chart axis lies to you
sh.getRange(tblStart + 1, 4, 3, 2).setNumberFormat('0%');

Create the spreadsheet with the Drive connector, then format it with
Apps Script, because the connector can create files but cannot edit cells.

Procedure:
1. Create the sheet, note its ID.
2. Write the Apps Script to a local .gs file first so we have it in version
   control and can re-run it after any rebuild.
3. Open script.google.com in my browser, create a project bound to nothing
   (standalone is fine, we openById), paste the code, save.
4. Run it and READ THE EXECUTION LOG back to me. Do not tell me it worked
   because it did not throw.
5. If I ask for a change later, edit the .gs file and re-run. Never
   hand-edit the sheet, or the next run wipes it.

Rules:
- Always make a hidden BACKUP tab before reorganising anything.
- The script must be idempotent: delete the sheet and rebuild it, so
  running twice is safe.
- Status colors use EXACT() so they are case sensitive.
- Put a real data validation dropdown on the status column.