# Learnings & Conclusions — Feb 3-5, 2026

## What We Did
- **Feb 3:** Bonanzo deck illustrations (Colin Ozawa style) — 14 slides, generated ~30 images, approved 4
- **Feb 4:** PHAT Foods brand strategy deck — full slide review, corrections, strategic core content
- **Feb 5:** Figma deck automation — built end-to-end workflow for templated slide creation, patched plugin, built 7 test slides

---

## 🔴 Problem 1: Pipeline Violations (The Biggest Issue)

**What happened:** Kitt repeatedly did creative work directly — generating images, writing prompts, reviewing quality, iterating on output. Even after creating explicit rules, the violations continued. Erica also bypassed the pipeline by generating and self-approving when sub-agent spawning failed.

**Root cause:** No hard technical enforcement. Rules in SOUL.md are just text — they can be overridden by urgency, context pressure, or "just this once" reasoning.

**Conclusion:** Pipeline discipline needs to be **enforced by architecture**, not willpower. If Kitt physically cannot run image generation, he won't. If Erica can only call Tatiana/Anton (not generate herself), she won't bypass.

**Skill/workflow to build:**
- `creative-pipeline-management` skill — Kitt's ONLY interface to creative work. It spawns Erica with locked-down instructions and returns only when Anton-approved output exists.
- Consider: remove nano-banana-pro from Kitt's direct tool access entirely.

---

## 🔴 Problem 2: Sub-Agent Spawning Breaks Inside Sub-Agents

**What happened:** Erica couldn't spawn Tatiana or Anton from within her session. She tried shell commands, direct API calls — nothing worked. This broke the entire pipeline because the coordinator couldn't coordinate.

**Root cause:** Technical limitation — sub-agents can't spawn sub-sub-agents reliably.

**Conclusion:** The pipeline must be flat, not nested. Kitt spawns ALL agents directly (Erica, Tatiana, Anton) and orchestrates the handoffs himself. Or: Erica gets a different mechanism to invoke team members.

**Skill/workflow to build:**
- Flat orchestration pattern: Kitt spawns Tatiana → gets output → spawns Anton with output → gets review → if rejected, loop → if approved, deliver.
- Erica becomes a planning/brief-writing agent, not a runtime coordinator.

---

## 🔴 Problem 3: Reporting Work Complete Without Verification

**What happened:** Multiple times, broken work was reported as done — black text on black backgrounds, overlapping slides, wrong content on wrong slides, files that didn't exist.

**Root cause:** "Task complete" messages were trusted without visual verification. Agents reported success based on command execution, not output quality.

**Conclusion:** Every deliverable needs a verification step built into the workflow — not optional, not "if you remember."

**Skill/workflow to build:**
- Mandatory screenshot/snapshot step after any visual output (Figma, image gen)
- Verification checklist that runs automatically before "done" status
- Anton's review should include actual visual inspection, not just prompt review

---

## 🟡 Problem 4: Infrastructure Before Execution

**What happened:** Feb 3 was chaotic (30 images, 4 approved = 13% hit rate). Feb 5 was productive (7 slides built correctly, workflow documented, templates registered). The difference: Feb 5 we built the infrastructure first.

**Root cause:** Jumping into execution without tooling/templates/workflow leads to thrashing.

**Conclusion:** For any new production task, spend the first session building the workflow, NOT producing output. Template everything possible. Document the exact commands. Then execute is just fill-in-the-blanks.

**Skill/workflow to build:**
- `deck-builder` skill pattern: template registry → content bank → automated build script → verify
- Reusable for any new deck project (not just PHAT Foods)

---

## 🟡 Problem 5: Style Consistency Across Generated Images

**What happened:** Each image generation was independent — no memory of previous approved images. Character appearance, background treatment, and style drifted across slides. Had to explicitly describe the "cyan hoodie boy" every time.

**Root cause:** Image generation models have no persistent state. Each prompt is a fresh start.

**Conclusion:** Style briefs need to be comprehensive, reusable documents — not ad-hoc prompts. Include: exact character description, approved color palette, background treatment, camera angle, lighting, things to avoid.

**Skill/workflow to build:**
- Style brief template (reusable across projects)
- Per-project style lock file: approved character sheet + palette + do/don't list
- Tatiana's job: maintain and enforce the style lock, not just generate from scratch each time

---

## 🟡 Problem 6: API Key / Infrastructure Fragility

**What happened:** GEMINI_API_KEY was invalid, memory_search broke, API 500 errors crashed agent sessions, Figma plugin needed patching.

**Root cause:** No health checks. Services break silently and we only discover during execution.

**Conclusion:** Pre-flight checks before any production run. Verify all dependencies work before starting the actual work.

**Skill/workflow to build:**
- Pre-flight check script: API keys valid? Services reachable? Plugin connected? Channel confirmed?
- Run automatically at session start or before any production workflow

---

## 🟢 What Worked Well

1. **Figma template-driven approach** — Components + instances + content bank = scalable, consistent, fast
2. **Content bank** (50-slide map with template assignments) — Separating content planning from production execution
3. **Thread discipline** (when followed) — Each slide in its own Discord thread kept feedback focused
4. **Workflow documentation** (figma-deck-workflow.md) — Once written, anyone can follow it
5. **Plugin patching** — Fixed real blockers (local components, image fills, base64) instead of working around them
6. **Competitive intelligence** (PHAT Foods) — Caught Nourish's Creamilux launch, Yali Bio pivot — real strategic value

---

## New Skills to Build

| Skill | Purpose | Priority |
|-------|---------|----------|
| `creative-pipeline-management` | Enforces delegation chain, prevents Kitt from doing creative work | 🔴 Critical |
| `deck-builder` (improved) | Template-driven deck production with content banks and auto-layout | 🟡 High |
| `style-brief-template` | Reusable style lock documents for consistent image generation | 🟡 High |
| `pre-flight-check` | Validates API keys, services, connections before production runs | 🟡 High |
| `flat-orchestration` | Pattern for Kitt to orchestrate multi-agent work without nesting | 🔴 Critical |
| `output-verification` | Mandatory visual verification step before reporting completion | 🟡 High |

---

## New Workflow Rules

1. **Infrastructure first.** New project? First session = build workflow + templates. Second session = produce.
2. **Flat orchestration.** Kitt spawns agents directly. No nested sub-agent spawning.
3. **Style locks.** Every visual project gets a style brief doc before any generation starts.
4. **Pre-flight always.** Check dependencies before starting production.
5. **Verify visually.** Screenshot/snapshot every visual deliverable. No "trust the command output."
6. **Content bank before production.** Plan all slides/pages/assets in a document first, then batch-produce.