Playbook · AI Experiments
AI Experiments
Diving deep into generative models, LLMs, and creative AI. It's less about perfect results and more about poking the bear to see what happens.
The 2026 Unlock
Skills × MCPs
"Skills + MCPs + understanding your own workflows is a crazy unlock."— me, after a year of finding out
Combo Lab
Pick hands. Pick a playbook. See what unlocks.
1 · The hands (MCP)
2 · The playbook (Skill)
n8n × Safe-Edit Standards
The day-job combo
The agent edits live automations the way a surgeon operates: copy the workflow first, pin test data, dry-run every write. Real business automations shipped, zero 3am incidents.
What the year beat into me
01
Hands without a playbook break things. A playbook without hands just watches.
MCP × Skill is a multiplication, not an addition. Either one alone is a party trick.
02
Cheapest model first.
A small model with explicit, step-by-step instructions beats a big one with vibes. Orchestrate, don’t hero.
03
Live systems get rituals, not confidence.
Pin test data, dry-run writes, checkpoint before anything risky. Confidence is not a rollback plan.
04
Local models crossed the line this year.
A free model on your own metal is now good enough to answer customers. The inference bill: almost nothing.
Prompt Library
A collection of curated prompts to jumpstart your experiments. Copy, paste, and modify for your own projects.
Filter by Category:
SQL-from-NL with Safety Guardrails
Natural-language to SQL that refuses silently-dangerous queries and always explains its assumptions. The assumption section is where you catch 90% of misinterpretations.
Context Engineering Checklist
A pre-flight checklist for agent briefs. The 2026 shift from prompting to orchestration means bad briefs compound — this catches them before they ship.
Multi-Agent Orchestrator Spec
Template for spec-ing a multi-agent system before writing code. Forces role clarity — most orchestration bugs are actually role-definition bugs.
Reflection Loop for Code Review
Two-pass prompt: the model reviews, then critiques its own review. Catches the ~30% of issues a single pass misses — most of which are style-cosplaying-as-substance.
Plan-Before-Edit Scaffold (Claude Code /plan)
Drop this into `.claude/commands/plan.md` to force a rigorous planning pass before any file edits. Costs 30 seconds, saves an hour of rework.
Deep Research Protocol
Structured research prompt that grades sources and surfaces disagreements instead of laundering them into false consensus.
RICE x MoSCoW Hybrid Prioritizer
Forces tradeoff conversations instead of wishlist politics. I run this before every cycle plan — it surfaces the items where data and instinct disagree.
Sora 2 / Veo 3 Storyboard Generator
Turns a rough idea into a shot-by-shot brief with camera, lighting, and continuity tags that 2026-gen video models actually respect.
Create A Blog Prompt
Prompt I used to create a blog using Google AI studio
More experiments and articles coming soon!
This space will feature case studies and deep dives into AI applications.