IPL GPT · T1 · 2026
The engine drafts. I decide what ships.
Over the last six months I have moved my production onto agentic workflows built in Claude Code. Skills, slash commands and subagents carry one job each, with named checks between the draft and the reader. Two engines carry this route. One produced checked cricket editorial around IPL GPT, a statistics product; the other builds the animated world of T1, a children's safety campaign. The framing and the final call on every piece stayed with me.
Skills · Commands · Gates
The cricket engine is a pipeline of single-purpose parts: three slash commands, a shelf of skills, and a gate between every stage. The commands stay few on purpose. Deciding what runs is a conversation, so each series plan lives in a brief file that the build routes on. The effort went into skills that emulate research-led data storytelling, and the parts that earn the polish are the checks.
GT vs RCB · 30 April 2026
One question, followed from brief to published card. Step the run.
01 · The /match-note brief
Would Ahmedabad deserve the broadcast's fortress framing once GT's home record was compared with RCB's record at the same venue?
“The host wins barely above a coin-flip rate.”
My call: challenge the familiar line without claiming this was the smallest home edge in the league.
02 · The product query
IPL GPT's checked cache supplied GT at 13-11 in Ahmedabad through IPL 2025, a 54.2 per cent win rate. RCB stood at 3-3 there, or 50.0 per cent.
“Three cache rows lifted into the card.”
Boundary: these venue figures came from IPL GPT. I selected the comparison and its editorial frame.
03 · The cross-validate audit
The GT total differed from one external table: 24 matches and 13 wins against 22 and 12. The two missing matches were the 2022 and 2023 finals at the venue, one won and one lost. Including those playoffs reconciled the denominator while leaving the win rate near 54 per cent. RCB's 3-3 venue record matched the external source exactly.
“Minor divergence, acceptable, reconciled.”
04 · The correction gate
The first draft carried the cache's through-2025 head-to-head figure of 3-3. The stale scope was caught before publication. Current scorecards confirmed RCB's win on 24 April 2026, so the published context became 4-3, with the venue rates explicitly scoped through IPL 2025.
“Caught by user pushback before publish.”
Final gate: I checked the correction and approved the scoped wording.
Carousel published · essay unpublished
Five measures moved together. This edition ran the full loop: writer and editor passes, engineered charts, a voice check and the four-pass design review before the slides were stamped. The eight-slide carousel condensed the career break for Instagram; the finished long-form version remained unpublished when the project closed.
What this route demonstrates
The product answered queries. The workflow kept source scope, external disagreement, corrections and publication status visible. I chose the reader's question, tested the claim and made the final call.
Children's safety campaign · 2026
T1 is a children's safety campaign built on one habit: a child names their trusted adult before they need one. The brand kit and campaign framework come from the client's team. The animated world is mine, and the work is visual first. The stack is a cast of six kids designed as locked characters, keyframe stills for every story beat, and animation run between the pinned frames.
I register each character in Google Flow, so every generated frame holds the same face, posture and shirt. A reel is planned as a storyboard of beats. I pin each beat as a keyframe pair, and the animation model works between the first and last frame; the current build runs this stack with Higgsfield's models. The first reel is assembled and delivered to the team. The campaign has not gone live, so its boards and cast sheets are not shown here yet.
