18 sessions of feedback + your SOP transcript, distilled into one skill upgrade, then proven on ajtozona.hu · 2026-07-20
18 sessions of feedback + your SOP transcript, distilled into one skill upgrade, then proven on ajtozona.hu
Master summary — the gist in 30 seconds
TL;DRWe mined all 18 prep sessions from the last 2 weeks, your SOP page and the dry-run transcript with Dani, and locked a plan: bake every lesson into the sales-call-prep skill, make it run faster by starting independent work immediately instead of waiting in line, and give it the 2-page Figma output you asked for — a workshop page with everything + speaker notes, and a clean client page the AI curates fully. Then we prove it on ajtozona.hu by critiquing the existing assets and regenerating only what fails.
Input: 2 weeks of your corrections, the SOP transcript, and 1,600 existing ajtozona assets. Output: an upgraded skill + a fresh 2-page ajtozona Figma board ready to present, with a before/after gallery showing exactly what improved.
Why this mattersEvery correction you made in those 18 sessions was you paying tuition. This plan is where the tuition turns into a machine that doesn't make those mistakes again — and finishes sooner.
flowchart LR
A["18 sessions<br/>+ SOP transcript"] --> B["Refined skill<br/>(rules + speed)"]
C["ajtozona assets<br/>(1,600 images)"] --> D["Vision critique"]
D --> E["Regen only<br/>what fails"]
B --> F["2-page Figma board<br/>ready to present"]
E --> F
1 · What gets baked into the skill
TL;DRTwelve locked rules — from '2× zoom, wider context on every evidence shot' to 'competitor ads are real screenshots, never re-creations' (that one was caught live on the dry-run).
Input: your repeated complaints and the SOP's canonical presentation order. Output: hard rules in SKILL.md that every future prep obeys automatically.
Why it mattersImage quality was your #1 recurring complaint, and the fabricated-screenshot moment was the scariest — both become impossible-by-default instead of remembered-by-luck. The presentation spine from your dry-run (competitor field → our ads → test tree → homepage teaser → ROI → calculator → guarantee → offer) becomes the board's fixed order.
2 · The new 2-page Figma output
TL;DRPage 1 'workshop': every asset and variant with a Hungarian sticky-note speaker line beside each. Page 2 'client': one AI-picked winner per family, clean, no internal numbers — final and presentable the moment the run ends.
Input: all generated assets + the call-insights research. Output: a light-loading Figma file where you present from page 2 and glance at page 1 for depth.
Why it mattersToday there is no speaker-notes surface at all, and curation was manual. You chose full AI curation — the run ends presentation-ready with zero assembly work from you.
TL;DRSame work, reordered: three research tracks start at minute zero instead of queueing, each asset family starts the moment its own data is ready, and the Figma build runs alongside the paperwork.
Input: the same pipeline stages. Output: the same deliverables, meaningfully sooner — like opening all supermarket checkouts instead of one line.
Why it mattersYou asked for speed without a token bill increase — reordering is free; duplication is not, so there is none.
flowchart LR
subgraph Before
a1["Scrape"] --> a2["Keywords"] --> a3["Ads intel"] --> a4["Assets"] --> a5["Figma"] --> a6["Run-doc"]
end
subgraph After
b1["Scrape"] --> b2["Keywords"]
b1 --> b3["Ads intel + PageSpeed<br/>start at T0"]
b2 --> b4["Each asset family starts<br/>when ITS data is ready"]
b3 --> b4
b4 --> b5["Figma ∥ Run-doc"]
end
4 · The ajtozona proof run
TL;DREvery existing asset goes through a vision critique against the new rules; only failures get regenerated. A before/after gallery shows old asset · verdict · new asset side by side.
Input: the existing ajtozona research (kept verbatim) and 1,600 images. Output: a new 2-page board + a scrollable gallery of exactly what the refined skill fixed.
Why it mattersIt tests every NEW part of the skill at a fraction of full-run cost, and the gallery doubles as proof the refinement actually moved quality.
⏭️ Next steps
TL;DRCopy the printed prompt into a fresh Claude Code chat — Instance 2 turns this handoff into a concrete checklist, then Instance 3 executes it autonomously.
You do one paste. Everything else — skill edits, critique, regen, Figma assembly — runs without mid-run questions; feedback happens afterwards via comment-html.
Why it mattersFresh context per instance keeps each stage sharp and cheap; the handoff bundle is written so nothing gets lost between them.
💡 Fun facts & practical stuff
TL;DRNuggets from the mining.
· The dry-run transcript shows the Figma file loading so slowly it nearly broke the session — 'keep the file light' is now a hard rule. · Your SERP now-vs-future zone was called 'pure gold' and survives untouched. · 18 sessions, 6 prospects, ~1,600 ajtozona images generated in one prior run. · The skill gains a new trick from your SOP: feed it a post-call transcript and it regenerates the flagged assets and logs the iteration. · Fast-mode ('assets 5 minutes before a call') is parked in the backlog by your call — it bit you twice, so it will come back as its own feature.
Why it mattersWorth knowing before the build starts.