# #brand-pipeline-spec **Kitt** (2026-03-24 06:42) # πŸ“Œ CE BRAND PIPELINE β€” LIVING SPEC **Last updated: 2026-03-24 | Status: Definition phase** This channel is the source of truth for everything we're building. Update it when decisions are made. Check it before starting any build work. --- ## WHAT THIS IS A **human-in-the-loop brand production suite**. It reduces the human time required to produce a complete, CE-quality brand from weeks of work to **6 structured decision moments totalling under 2 hours**. The human is not patching a broken system β€” they are the taste layer, positioned at the moments where taste genuinely belongs. The system handles everything else. ## WHAT THIS IS NOT - Not an autonomous brand generator (AI doesn't have taste, and we're not pretending it does) - Not a replacement for CE's creative judgment β€” it compresses the time that judgment requires - Not a product someone else can run without Assaf yet (that's Phase 4, conditional on calibration data) ## THE HONEST CEILING > "This doesn't give AI taste. It gives AI a model of CE's taste, trained on real decisions over time." **Kitt** (2026-03-24 06:42) Everything we build should be consistent with that sentence. **Kitt** (2026-03-24 06:42) ## THE 8 TOOLS + 6 HUMAN TOUCHPOINTS The pipeline runs autonomously between touchpoints. Humans appear at 6 designed moments β€” with the right information, for minimum time, making the decisions only they can make. ``` TOOL RUNS AUTONOMOUSLY HUMAN TOUCHPOINT ──────────────────────────────────────────────────────────────────────── 01. Brief Enricher β†’ research + gap-fill β†’ T1: Review enriched brief 02. Research Engine β†’ market / comp / audience β†’ (feeds T1, no extra gate) 03. Strategy Maker β†’ 3 positioning options β†’ T2: Pick one + say why 04. Territory Finderβ†’ 3 visual directions β†’ T3: Pick one + say why 05. Style Forge β†’ style guide + JSON β†’ T4: Approve or adjust 06. Asset Generator β†’ image batches β†’ T5: Select the cuts 07. Deck Builder β†’ Google Slides deck β†’ T6: Review narrative flow 08. Page + Social β†’ web page + posts β†’ T6: Review before publish ``` **Kitt** (2026-03-24 06:42) **Time at each touchpoint:** 10-15 min max **Total human time per brand:** under 2 hours **Total elapsed time:** 3-4 days (autonomous work runs between touchpoints) --- ### TOOL 01 β€” Brief Enricher - **Input:** anything β€” even one sentence - **Does:** fills gaps through research. Sector context, category norms, competitors, audience signals - **T1:** Read enrichment. Correct what's wrong. Add what research couldn't know (internal constraints, relationships, fears). 10 min. - **Output:** validated brief that feeds all downstream tools ### TOOL 02 β€” Research Engine *(runs inside Tool 01)* - Julia runs market/competitive/audience research - Output populates the brief automatically β€” no separate human stop ### TOOL 03 β€” Strategy Maker **Kitt** (2026-03-24 06:42) - **Input:** enriched brief - **Does:** produces 3 positioning directions. Each: named angle + 3 reasons viable + what it rules out - **T2:** Read 3 options. Pick one. Write one sentence about why. Sentence is logged. - **Why 3:** not good vs. bad β€” 3 defensible directions. Human taste determines which is most CE. - **Output:** positioning.json β€” chosen angle, rationale, constraints locked ### TOOL 04 β€” Territory Finder - **Input:** positioning.json + taste board query - **Does:** queries the signal library for what's emerging in this category. Produces 3 named visual territories β€” each with: typography pair, color temperature, photography style, 3 real brand references, rationale - **T3:** Sees 3 territories as visual cards (not text). Picks one. Can flag elements from others to carry over. Writes one sentence. - **⚠️ THIS IS THE TASTE GATE.** Everything downstream follows from this pick. - **Output:** visual-direction.json β€” chosen territory locked with reference images ### TOOL 05 β€” Style Forge - **Input:** visual-direction.json - **Does:** produces VISUAL-STYLE.md + visual-style.json + Gemini prompt framework. Scores itself against 3 taste board references before human sees it. - **T4:** Review style guide against references. One correction maximum. Not a full revision. **Kitt** (2026-03-24 06:42) - **Output:** visual-style.json (locked) β€” feeds all downstream tools ### TOOL 06 β€” Asset Generator - **Input:** visual-style.json - **Does:** generates image batches. Vision model runs criteria-check (technical compliance) before human sees any image. Removes failures. - **T5:** Selection pass. Human sees only the batch that cleared QA. Picks 3-5 per use case. - **Output:** asset-library/ β€” curated image set, tagged by use case ### TOOL 07 β€” Deck Builder - **Input:** positioning + style guide + asset library - **Does:** builds Google Slides deck. CE design system applied automatically. Copy written from positioning doc. - **T6 (combined with Tool 08):** Review deck + page + social in one pass. Final review before delivery. - **Output:** complete deliverable set **Kitt** (2026-03-24 06:42) THE SHARED DATA MODEL Everything flows through one evolving object: the **Brand Brief**. Starts sparse, gets richer at every stage. Each tool reads from it and writes back to it. ``` brand-brief/ β”œβ”€β”€ intake/ ← what the client provided β”œβ”€β”€ enriched/ ← what research added β”œβ”€β”€ positioning/ ← chosen strategic direction + rationale β”œβ”€β”€ visual-direction/ ← chosen design territory + rationale β”œβ”€β”€ style-guide/ ← visual-style.json (locked) β”œβ”€β”€ asset-library/ ← curated image set β”œβ”€β”€ deliverables/ ← deck, page, social └── decisions/ ← log of every human choice + context ← CALIBRATION CORPUS ``` The `decisions/` folder is how the system learns over time. Every gate decision logs: what were the options, what was chosen, what rationale was given. After 10-15 runs, patterns emerge and options start pre-ranking toward what CE historically selects. **Kitt** (2026-03-24 06:42) ─── THE TASTE LAYER Taste isn't replaced. It's compressed. Each touchpoint is designed to require the minimum decision for maximum downstream accuracy. ``` | What taste does | Where | Tool | | ------------------------------- | ----- | ------------------------------------------- | | "This positioning is more CE" | T2 | Strategy Maker β€” 3 options side by side | | "This visual world is right" | T3 | Territory Finder β€” visual cards, not text | | "These criteria need adjusting" | T4 | Style Forge β€” one correction, not a rebuild | | "This image is CE / this isn't" | T5 | Asset Generator β€” selection interface | ``` ─── THE TASTE BOARD AS INFRASTRUCTURE **Kitt** (2026-03-24 06:42) Not a mood board. A **reference signal database** that Tool 04 queries. Each entry tagged with: β€’ **category** β€” what sector (food, fintech, personal brand, etc.) β€’ **timing** β€” emerging / at peak / declining β€’ **design_signals** β€” color temperature, typography weight, photography style, mood β€’ **ce_resonance** β€” high / medium / low (requires human taste to maintain β€” intentional) β€’ **notes** β€” what Assaf saw in it When Tool 04 runs for a food brand: queries `category: food AND timing: emerging AND ce_resonance: high`. Territories generated are grounded in what CE actually finds interesting now. **The CE resonance field stays human.** Assaf curates the board. That curation IS the taste input. The system amplifies it. ─── THE EYES WORKAROUND (Vision QA) **Kitt** (2026-03-24 06:42) AI is blind by default but not by necessity. At Stage 5 (Asset Generator), vision model evaluates images against explicit style guide criteria before human sees them. Instead of "does this look good?" β†’ "does this image exhibit [specific characteristics from the style guide]?" β€’ Warm golden light present? βœ“/βœ— β€’ Dark shadows visible? βœ“/βœ— β€’ Texture detail sharp? βœ“/βœ— Score = % of criteria met. Below threshold β†’ loop back to generation with specific critique. Human only sees images that cleared technical QA. **This is a floor, not a ceiling.** It removes garbage. Taste still lives at the gate. **Kitt** (2026-03-24 06:43) BUILD PLAN PHASE 0 β€” Design the foundations _(before any code, ~1 week)_ Design sessions only. No building until these are done. β€’ [ ] **Design visual-style.json schema** β€” the connective tissue between Stage 3, 4, 5, 7, 8. Validate against Phat (real example). This comes before everything else. β€’ [ ] **Design taste board tagging schema** β€” define the controlled vocabulary for all tag fields β€’ [ ] **Map one signal-to-decision chain forensically** β€” take Phat, reconstruct: what was the actual input signal β†’ what was the actual design decision β†’ what was the inferential step between them. Write it down explicitly. If the step can be made explicit, the framework is real. β€’ [ ] **Write the "good brief" definition** β€” what makes a brief good vs. merely complete. This informs Tool 01. **Gate:** Don't start Phase 1 until all four Phase 0 artifacts exist. ─── SPRINT 1 β€” Harden the connective tissue _(2-3 weeks)_ **Kitt** (2026-03-24 06:43) β€’ [ ] Build visual-style.json into the pipeline (Stage 4β†’5β†’7β†’8) β€’ [ ] Write visual-territories SKILL.md (Stage 3 as a proper skill, from Phat pattern) β€’ [ ] Implement taste board tagging on existing candidates (50 entries minimum) β€’ [ ] Build Stage 0 brief intake β€” completeness + quality check β€’ [ ] **Test: semi-cold brief** (real category, minimal CE history β€” NOT Mission). Log every gate decision. **Success:** One complete run, no broken handoffs, outputs coherent with each other. ─── SPRINT 2 β€” The quality layer _(3 weeks)_ β€’ [ ] Build signal-to-decision framework as a skill (from Phase 0 forensic work) β€’ [ ] Implement vision model QA at Stage 5 (criteria-based pre-filter) β€’ [ ] Build decision capture protocol β€” every T2, T3, T4, T5 logged with context β€’ [ ] Stage 4 style guide gate β€” scored against taste board references β€’ [ ] **Test: second semi-cold brief**. After run: Assaf + Anton both score outputs at each gate. Document where they agree/diverge. This is the calibration. **Kitt** (2026-03-24 06:43) **Success:** Assaf reviews at 6 touchpoints. Anton handles quality between them. No silent drift. ─── SPRINT 3 β€” Full pipeline, hardened _(2 weeks)_ β€’ [ ] Tool 07: Deck Builder (harden from Phat scripts β€” reusable, not one-off) β€’ [ ] Tool 08: Page + Social connected through visual-style.json β€’ [ ] Make taste board live β€” automated signal ingestion from trend-scouting β€’ [ ] Full end-to-end test: cold brief, full pipeline, all 6 touchpoints β€’ [ ] Calibration review: how close is Anton to Assaf after 3 runs? ─── PHASE 4 β€” Productize _(after Sprint 3 stable β€” conditional)_ β€’ [ ] Brief intake UI (client-facing form) **Kitt** (2026-03-24 06:43) β€’ [ ] Progress dashboard (where in the pipeline, what needs human input) β€’ [ ] Deliverable packaging (standard output set per run) β€’ [ ] Pricing/packaging model β€’ [ ] **Gate:** "Quality without Assaf" only viable once decision corpus is validated β€” minimum 10 full runs with calibration data showing Anton scores correlate with Assaf's at a meaningful rate. Do not promise this to clients before that gate. ─── OPEN DECISIONS (update as resolved) ``` | # | Decision | Status | Owner | | --- | --------------------------------------------------- | ------ | ------- | | 1 | What does a "good brief" look like vs. complete? | ⬜ OPEN | Phase 0 | | 2 | visual-style.json schema fields | ⬜ OPEN | Phase 0 | | 3 | Taste board tagging vocabulary (controlled fields) | ⬜ OPEN | Phase 0 | | 4 | Which semi-cold brief to use for Sprint 1 test? | ⬜ OPEN | Assaf | ``` **Kitt** (2026-03-24 06:43) ``` | 5 | What's the Phase 4 product price point / packaging? | ⬜ OPEN | Later | | 6 | Who operates this besides Assaf? | ⬜ OPEN | Later | ``` ─── CONTEXT LOG _(Append decisions and key moments here as we build)_ **2026-03-24** β€” Product definition session complete. Three rounds of Anton critique. Decisions made: β€’ This is a human-in-the-loop product, not autonomous β€’ Minimal intervention = time saved + decision optimization β€’ Quality ownership = Assaf for now, with calibration path toward Phase 4 β€’ Taste board is the reference signal infrastructure, not a mood board β€’ Phase 4 is conditional on calibration data β€” not promised **Kitt** (2026-03-24 06:48) ## 🧡 PHASE 0 β€” Design the Foundations **Status: ACTIVE | Gate: don't start Sprint 1 until all 4 artifacts exist** Phase 0 is design-only. No code. No building. We're designing the schemas and frameworks that everything else is built on. Get these wrong and we rebuild everything downstream. **The 4 deliverables:** A. visual-style.json schema β€” connective tissue between Tools 4β†’5β†’6β†’7β†’8 B. Taste board tagging schema β€” controlled vocabulary for the signal library C. Phat signal-to-decision forensic map β€” prove the framework has real content D. "Good brief" definition β€” what makes a brief good vs. merely complete Starting with A. Reply in thread.