Summary — 2026-05-24

Third pass — 2026-05-24 (Taste Pipeline — references-to-skill)

One item this pass. Surfaced by Anthony from a third-party post describing a reproducible pipeline that turns a curated reference set into an enforceable design skill. Direct counterpart to what Dylan currently hand-curates as the prompt anchor library, and direct mitigation of today's twin visual-misread fails.

The pipeline (six steps, as described)

1. Curate references. High-resolution, tightly cropped on the details that make the reference work. Load-bearing input — "if the references suck, so does the skill."
2. Dual-model per-image analysis. Each reference analysed independently by two frontier models (Opus 4.7 + GPT 5.5 in the source post). Focus is on why the reference works as design — layout, spacing, typography, rhythm, composition, hierarchy. NOT functional / what-it-does analysis. Output: reference-NN/opus-4-7-analysis.md + reference-NN/gpt-5-5-analysis.md.
3. Anonymised fusion. Strip model attribution from the two analyses, hand both to GPT 5.5, get fused-analysis.md per reference. Anonymisation prevents self-preference bias when the fusing model recognises its own output.
4. Chunked synthesis. Group fused analyses into chunks of ~6–8. Each chunk fuses into a chunk-level rule synthesis. Key insight: all-at-once synthesis summarises instead of preserving granularity. Chunking preserves rules.
5. Final rule-set. One model pass fuses chunk syntheses into a single rule-set markdown.
6. Skill. Concrete imperative instructions extracted from the rule-set. Enforces constraints. No vague taste words.

Why P=3 (personal productivity)

Dylan currently hand-curates Design/PromptLibrary/<brand>/anchors.md per the SOP in Design/PromptLibrary/README.md. Five blocks (voice anchor · style anchors · character locks · composition + lighting directives · banned moves) at ~2–4 hrs senior-CD time per brand. Brands live in the library now: Articulate · Articulate-OperatorSequence · Kitroom · Roche-Debbie · _template. Every client engagement seeds a new one. Roger Deakins persona faces the same problem on cinematography references. Pipeline replaces hand-curation with reproducible synthesis — brand seeded in one session, not three.

Why Q=3 (quality)

Three quality moves we don't currently run anywhere in our stack:

• Dual-model debias. Single-model analysis carries that model's biases (e.g. today's Opus-led FAIL-VISUAL-MISREAD-OGILVY-MARK — sans-vs-serif misread). Two models triangulate.
• Anonymised fusion. Stripping attribution before the fuser sees the input removes self-preference bias. Cheap, structural, we don't do it.
• Chunked synthesis preserves granularity. Generalises beyond design — applies to brand-voice synthesis from N transcripts, failures.md weekly digest, any "synthesise N things into rules" pipeline we run. The chunking insight is the most portable piece of the writeup.

Direct mitigation of today's twin fails (FAIL-VISUAL-MISREAD-OGILVY-MARK + FAIL-WORDMARK-PROXY-MISCALIBRATED) — both single-model misreads of visual references that a dual-model triangulation likely catches.

Why S=2 (sellable, not 3)

Maps onto a real Articulate offer surface: "Brand Skill from Your References" — take a client's mood board / brand book / existing creative → produce an enforceable image-gen skill scoped to their brand. Productisable add-on to the brand-voice plugin pattern already in our stack (brand-voice covers copy; this covers visual). Slots into MEP brand-discovery phase. But it is not a standalone product — it is a layer inside Dylan's workflow. S=2 holds until first client engagement scopes "build us a visual brand skill from these references" as a discrete line item, at which point re-score to S=3.

Class-match against failures.md

• 2026-05-24 FAIL-VISUAL-MISREAD-OGILVY-MARK (this morning) — single-model misread, propagated through downstream artifacts. Dual-model pipeline likely catches it at step 2.
• 2026-05-24 FAIL-WORDMARK-PROXY-MISCALIBRATED (this morning) — execution within correct category was wrong because no parity-before-comparison gate. Pipeline forces explicit rules (weight, scale, proxy) extracted from refs.
• 2026-05-18 Takis four-brand-cycle (brand-spec drift) — pipeline reduces drift by anchoring to explicit, deterministically-generated rules instead of vibe / persona memory.
• 2026-05-20 mymentor carousel (wrong-tool fail, hero-vs-editorial) — pipeline establishes "fit-for-purpose" criteria as explicit rules up front.

Four class-matches, all in the same axis (visual rules unenforced because no structural source of truth). Strong adopt signal.

Re-score triggers

• (a) Pilot fails — if the dual-model + chunked-synthesis pipeline applied to Articulate's own brand produces anchors that are worse than the hand-curated baseline, Q drops to 2 (total 7/9 — still act-on but as enhancement, not replacement).
• (b) First paid client engagement scopes the pipeline as a discrete deliverable — S promotes to 3 (total 9/9).
• (c) Chunking insight generalises — if applied to brand-voice synthesis or failures.md digest with measurable quality lift, score a second item separately ("chunked-synthesis-as-pattern").
• (d) A third frontier model joins the analysis stage (Gemini 2.5 / Mistral Large / etc.) — re-evaluate dual-vs-triple-model triangulation cost/quality.

Build path (recommended option B — see decisions.md)

Scaffold as SuperSebastian/playbooks/taste-to-skill.md and extend Dylan's /dylan prompts <brand> subcommand with a --from-references flag that fires the pipeline. Not built this pass — playbook scaffold is the next move, pilot is the trigger after that.