Illustrative diagram · five phases that can be chained under a single command; the CAPTURE phase is what makes the skill self-improving.
🎷 What is a mega-skill
Most skills solve a task: generate an itinerary, review a workflow, audit a website. A mega-skill does something else—it weaves together multiple phases of complex work under a single entry point. You say one sentence, and the skill guides research, proposal, brand, and operations like a pipeline.
🎯 The definition
Improvised Intelligence defines itself as a unified engine that takes a business from "cold prospect" to "fully operationalized client"—with workflow capture at every step. It orchestrates four interconnected phases and a fifth that observes everything:
- •PROSPECT — competitive intelligence and market research on the target.
- •PITCH — a personalized proposal with 5 deliverables that close the deal.
- •BRAND — guided creation of a brand book in 8 phases.
- •OPERATE — a team of AI agents that runs ongoing operations.
- •CAPTURE — detects repeatable patterns and generates new skills from completed work.
💡 Practical tip
The mega-skill doesn’t start out big on purpose. It grows when you realize that several of your skills always run in the same order. Instead of calling three skills in sequence every time, you combine them under one command and a phase detection table. Size isn’t the goal—eliminating manual coordination is.
One command goes in, the skill chooses the path.
Run sequentially or independently.
The request selects the phase.
The workflow becomes a new skill.
🧠 Benefit 1: replaces an agency workflow
The first benefit is the most visible: the skill replaces an entire agency workflow — research, pitch, brand strategy, and AI operations — in a single chain of commands. What would normally pass through several people and several tools happens through simple phrases.
| You say | What the skill produces |
|---|---|
| pitch [company] | Researches the business, finds competitors, builds a personalized landing page, 3 PDFs, and a video script. |
| build a brand book for [client] | Guides you through 8 phases and produces a complete brand identity system. |
| build my agent system | Architects a multi-agent team with memory, escalation rules, and a COO that audits itself. |
✓ Why this is strong
- ✓One request triggers a pipeline of deliverables that work together.
- ✓The phase 1 research feeds into the phase 2 pitch—nothing gets lost.
- ✓The phase 3 brand becomes the voice foundation for phase 4 operations.
✗ What to avoid when copying
- ✗Try to generate all 5 deliverables without doing the research first.
- ✗Chaining phases that don't share context just creates a task menu.
- ✗Use generic placeholders instead of the client’s actual business.
Target dossier.
5 deliverables.
8 guided phases.
Agent team.
📈 Benefit 2: gets smarter with every session
The second benefit is the most interesting for anyone studying skill architecture: Improvised Intelligence gets smarter with every session. A built-in detector — the Workflow Spotter — captures what you do and turns it into reusable skills.
🔎 When the Spotter triggers
It evaluates "is this worth turning into a skill?" whenever ANY of these happens:
- • 10+ back and forth in a single project
- • 5+ tool calls for the same result
- • 3+ output files created
- • A complete Research → Analysis → Creation → Delivery arc
- • The user says “let’s do this again” or “for next time”
The detection checklist
Before suggesting a skill, the skill asks itself—if 3+ are checked, it offers:
□ Was it more than a simple Q&A?
□ Did it involve 3+ distinct phases?
□ Were reusable artifacts created?
□ Would the user probably do something similar again?
□ Could 50%+ of the work be systematized?
□ Did I discover something the "future Claude" shouldn't have to rediscover?
💡 Practical tip
This is the same self-improvement principle you'll see in Taproot (6.2) and on the /sessionend (6.3): capture the work instead of letting it evaporate. Steal this idea for any skill of yours — a block at the end that asks "does this keep happening? want me to save it?".
What you do.
Repeatable pattern.
Turn it into a skill.
SKILL.md ready.
⚙️ Benefit 3: born in production
The third benefit is a quality guarantee: every piece of the skill runs on real business fronts now — was built in production, not in theory. That changes everything about the quality of the defaults, brand rules, and description.
High-value patterns the skill already catalogs
⚠️ Attention: production skills have production rules
Skills born in production carry explicit rules that you wouldn’t come up with on paper: a sensitive information protocol (never automatically learn financial, legal, or health data), channel limits, a locked brand voice. When writing your own, follow this discipline—clear constraints are more valuable than vague freedom.
Quality standard before delivering any output
- Specificity — refers to the real business, not placeholders?
- Research-backed — are the claims based on the recon data?
- Actionable — does each section end with a clear next step?
- On-brand — does it sound like the brand, not generic consulting?
- Complete — does it include every item on the phase checklist?
Not on the slide.
Sensitive data protected.
Brand consistency.
Checklist by phase.
🗺️ Phase Detection: The Command Chooses the Path
How can such a large skill have a single entry point without turning into a maze? With a phase detection table. The skill reads the request, matches it to the right phase, and routes itself. If it’s ambiguous, it asks one disambiguation question — not an interrogation.
Read the request
"pitch X" → PROSPECT+PITCH · "brand book for Y" → BRAND · "build agent system" → OPERATE · "onboard Z" → all in sequence.
Disambiguates with a question
If you can’t decide: "Are we prospecting, onboarding, or building operations?" Just one. No 10-item menu.
Runs the phase (or chain)
Each phase can run on its own or in a sequence. At the end, the CAPTURE phase evaluates whether what happened is worth turning into a skill.
💡 Reusable pattern
The “phase detection table” is one of the most useful tricks for any skill with modes. Instead of a confusing mental if-else, you give Claude a table Se o usuário diz → Fase → Ação. It matches the request and takes action. Copy this table into your next multi-mode skill.
Request → phase.
Only if ambiguous.
The phase runs on its own.
Or all together.
🔁 How to Use It in Practice
In practice, you trigger the skill with command phrases. Each one starts the right phase; together, they run the entire chain. Below are copyable prompts to try—adapted for general use, without tying them to any specific brand.
pitch [nome do negócio] Pesquise o negócio (avaliação, concorrentes, o que os clientes procuram), encontre a "lane não reivindicada" dele e monte: uma landing personalizada, 3 PDFs (calendário de conteúdo, posicionamento, snapshot de mercado) e um roteiro de vídeo curto.
build brand book for [cliente] Conduza as 8 fases guiadas. NÃO pule fases nem assuma respostas: descoberta → missão → promessa → taglines → voz e tom → regras de vocabulário → bancos de frase por avatar → compilar. Confirme comigo a cada fase antes de seguir para a próxima.
capture this workflow Run the detection checklist on what we just did. If 3+ items match, propose a new skill: suggested name, what it would include, and the trigger phrases. Don’t build it yet — just show me the proposal.
📤 Output example
Illustrative recreation of the block that Workflow Spotter prints when it detects a pattern (not a real screenshot):
💡 Personalize before using it for real
For real-world use, replace the brand voice and defaults with the your identity (PDF colors, signature phrase, tone). The mega-skill was designed to coexist with your philosophy—not impose the original author's.
🏋️ Hands-on exercises
Map your phases
Choose a task of yours that always follows the same order. List the phases (3 to 5) and write the detection table: "If I say X → phase Y → action Z".
Write the capture checklist
Adapt the Workflow Spotter's 6-question checklist to your domain. Set the threshold (how many items need to match to suggest a skill).
Create a runnable SKILL.md (mini mega-skill)
Write a SKILL.md in two phases under a single command, with a detection table and a capture block at the end. Skeleton:
Save in ~/.claude/skills/meu-pipeline/SKILL.md, open Claude Code and test with "run acme".
📌 Module summary
Next module:
6.2 — 🧬 Taproot: skill tree (meta-skill)