RULEBOOK · AUTOMATION · SHORT-FORM 2025 — 2026
WORK. I turn rejections into rules, rules into automated checks, and automate the production itself.
The script-naturalness study behind these rules (judge evaluation, national corpus) is on Research ↗ .
GUIDELINE · QA
16 rules · 0 violations
Rejection reasons → rulebook → automated checks
EDIT AUTOMATION
93 s → 26.8 s
Render time after changing the capture method, identical output hash
JOJO INVESTMENT GROUP
5–6 a day
AI-avatar shorts, script to final edit
JOJO INVESTMENT GROUP
700–800 / month
Team target metric (monthly new-member inflow) — I owned script and edit
01 / GUIDELINE · QA Pagard AC
One rejection → one rule → one automated check Pagard AC · Aug 2026 – present · short-form & chart video automation (Claude Code · Python) · final sign-off by the team lead
RULEBOOK · WHAT EACH RULE RECORDS
MACHINE-JUDGED
Safe-area overflow, clipped labels → automated checks; full re-scan of every scene: 0 violations
HUMAN-JUDGED
Awkward staging → frame-by-frame analysis of approved finals; assemble only from directions actually used
16 rules 37 confirmed constraints (at the time)
Result: no rejection came back a second time
ACTUAL RULEBOOK ENTRY
# brand/EDIT-RULEBOOK.md — ⑮ 안전 영역
반려: 글자가 안전 영역 밖으로 나감
규칙: layout padTop 216 / padBottom 162 고정
검사: probe-labels 전 장면 재검사 → 위반 0QA · WIDENING WHAT GETS CHECKED
Character-level script ↔ subtitle diff → 12 typos and 1 missing sentence found Subtitle sync checked against word timestamps → 0 drifting cues Measured 13 shorts: target 45 s, median 55.9 s → reset the script-length standard Dropped Korean auto spell-checkers from the pipeline — too many over-corrections 02 / AUTOMATION Pagard AC
Automating the edit itself Pagard AC · Aug 2026 – present · driving Premiere and After Effects with Claude Code agents
ROLES · SHARED MEMORY
Lead — review · log · merge
Confirmed Premiere had no external automation path → used Photoshop as the bridge to drive Premiere and After Effects Delivered layered After Effects files instead of flat mp4s, so every edit stays fixable Changed the capture method: render 93 s → 26.8 s with an identical output hash One SQLite file holds the work log — a fresh session picks up the context Each script scene is auto-built from a matching segment of real MT5 market data Auto-generated After Effects composition — 17 layers Edit file: change text or color in Premiere and the chart scene follows 03 / TEST DATA · DECOYS Pagard AC
Testing with decoys, not just answers Pagard AC · Sep 2026 – present · rules that find the MT5 price segment matching each script scene (MQL5 · Python)
Synthetic chart
one real pattern
→ + Decoys
replays of past real misses
→ Rule
passes only on the real one
10 regression tests — a case type missed once stays in as a decoy Prices × 1,000 gave identical scores → the rule isn't fitted to a price range MT5 connected read-only, no order functions · each pick logged as a storyboard EVALUATION PRINCIPLE
Test not only whether it gets the answer, but whether it avoids the look-alikes. Misses aren't deleted — they become the next test's decoys.
04 / SHORT-FORM JOJO INVESTMENT GROUP
Short-form judged by sign-ups, revised every week Jojo Investment Group, video team · Oct 2025 – Feb 2026 · AI-avatar stock shorts, no filming
PIPELINE · SAME HOST, FACE, VOICE EVERY DAY
→ → Nanobanana PRO
wardrobe · styling
→ → AE · Premiere
captions · cuts · finish
LOOP · EVERY WEEK
→ Response
new member sign-ups
→ Weekly review
pacing · tone · caption density
5–6 shorts a day, script to final edit, working against the team's target metric (700–800 new members a month) Catalogued where ElevenLabs lost intonation (numbers, units, loanwords) → a script-stage sentence-splitting rule that cut rework Ran the company YouTube channel's uploads, titles, and thumbnails Channel strategy and script direction were team calls; script, generation, and editing were mine 05 / FACE SWAP PERSONAL R&D
Putting a chosen face into an existing shot Face-swap inpainting workflow · ComfyUI (InstantID + ControlNet + Depth)
Core nodes — InstantID (identity) · Zoe-Depth + ControlNet (angle and lighting) · KSampler (dpm_adaptive, denoise 0.40). Input and result image panels are cropped out.
MASK
Only the face area is masked; hair, hands, and background stay original.
HOW
Took a public 18-node face-inpainting workflow and extended it to 22 nodes — keep an existing shot's composition and lighting, regenerate only the face area InstantID holds the identity; ControlNet + Zoe-Depth keep the original angle and lighting; only the masked area is regenerated Compared 7 samplers on one seed → picked dpm_adaptive, which tracked the reference texture best (judged on eyes and hair) Collected identity-preserving prompts from Reddit and communities, adapted to my use USE
Keep a proven shot's composition and lighting, swap only the person — fast variations on one concept.
CONTRIBUTION 01 · 02 rulebook and automation: designed and run by me, final sign-off by the team lead · 03 decoy tests: solo · 05 face swap: extension of a public workflow, solo · 04 Jojo Investment Group: script, generation, and editing by me; channel strategy was a team call
