POL FERRANDO /
[ 01 / INTRO ]
POL FERRANDO
CREATIVE TECHNOLOGIST / DEVELOPER
POL
FERRANDO
CREATIVE
TECHNOLOGIST
DEVELOPER
FOCUS · 2026

I'VE FOCUSED ON LEARNING HOW TO USE AI TO EXTEND MY CREATIVITY.

SEE IT IN ACTION
[ 02 / DEMO ]

AI adapted
to your workflows.

Pick a prompt. Hit Send. See what Claude returns when it already knows your projects, your decisions, and how you work.

Claude, anything I should handle before I start?
Claude

[Client A]'s last sprint is ready to bill, €X,XXX from your day-rate formula, saved as [Client A]-2026-04.docx. [Client B]'s invoice #47 is 18 days overdue, chaser drafted in your Gmail drafts, your usual tone. Your accountant's Q4 ask: Books.xlsx is filled.

One thing I didn't touch: the [Client C] contract renewal. Scope changed, that's your read before I draft.

[ Layer 1 of 4 · CLAUDE.md ]

Every demo above is Claude Code running inside four user-authored layers. Miss one and the system collapses back to stateless chat. Every business needs all four, tuned to how that business actually works.

CLAUDE.md

The memory layer

A plain-text file Claude reads on every tool call. Principles, routing, what to file where, what to check first. It's where the durable shape of your work lives. It scopes from global to project, and when a topic grows past a few lines it breaks out into its own file under .claude/rules/ — same always-on behavior, one chapter per file, path-scoped so nothing clutters unrelated work.

~/.claude/CLAUDE.md Global

Applies to every session on your machine. Personal style, safety defaults, cross-project preferences.

Never use --no-verify to skip hooks.
Ask before destructive ops (rm -rf,
  force push, dependency removal).
Default to the naive option.
No em dashes in prose or commits.
./CLAUDE.md Project

Applies inside a specific project. Tech stack, conventions, where narrative vs implementation live.

Tailwind only. No custom CSS files.
Copy lives in Cortex. Don't rewrite
  pitch copy here; edit the spec.
Budget: 300KB gzipped JS.
.claude/rules/*.md Modular, one topic per file
  • journaling.md Every file-modifying session writes a journal entry per project, with a defined format and cadence.
  • tool-routing.md Structured task queries go through the Obsidian MCP via Dataview. Not Glob + Read loops.
  • pushback.md When to challenge instead of validating. Named failure modes. Hold position under pressure unless new evidence lands.
  • self-improvement.md Triggers for proposing edits to the instruction layer itself. The system learns from its own mistakes.
// from pushback.md
Verdict first, then reasoning.
"That won't scale because Y."
Not "Interesting, but have you considered..."

Consequence: you stop re-explaining your standards every session. The next one starts already knowing them. Consistent behavior across projects, across topic shifts, across a year of work.

[ Layer 2 of 4 · Skills ]

Skills

The procedures

Reusable workflows in .claude/skills/. Each has a "Use when X" description. Claude matches your prompt against those descriptions and loads only what fires. Roughly an order of magnitude cheaper than equivalent MCPs because nothing runs until it triggers.

  • session-start Orient from recent journals. Reads per-project state, self-heals gaps from transcripts.
  • capture-and-record Save URLs, observations, testable beliefs the moment they surface. No prompting required.
  • log-decision Record a choice between alternatives with the reasoning. Retrievable six months later.
  • fanout-synthesis Spawn 2-5 parallel subagents on distinct slices of a landscape, then synthesize one report.
  • conduct-review Systematic evaluation of an internal artifact. Findings and recommendations, not execution.

Consequence: procedures stop needing re-derivation. The right steps run the right way, every time the trigger fires.

[ Layer 3 of 4 · MCP servers ]

MCP servers

The connectors

External systems Claude calls through. Each MCP server exposes typed tools Claude invokes inside the conversation. Gmail to draft mail. Calendar to schedule. Obsidian to query your vault. Chrome DevTools to drive a real browser.

  • obsidian Dataview queries over your notes, semantic search, patch frontmatter without re-reading the file.
  • gmail Draft replies, scan threads, manage labels. The email is the tool, not a copy-paste target.
  • calendar · drive Schedule work against real availability. Pull source documents Claude can read directly.
  • chrome-devtools Drive a browser, run Lighthouse, inspect live pages. Test your own site while building it.
// actual call Claude makes, mid-conversation
search_vault({
  queryType: "dataview",
  query: "TABLE status FROM 'Projects'
          WHERE type = 'task' AND priority = 'A'
          AND !completed"
})

Consequence: Claude acts on your systems, not just talks about them. The operation is the conversation.

[ Layer 4 of 4 · Agents ]

Agents

The specialists

Sub-agents with isolated context. Main Claude delegates; results come back without polluting the primary conversation. Each agent has its own tool whitelist and its own model tier. You can spawn several in parallel.

  • Explore Fast codebase surveys, pattern searches, open-ended questions across many files.
  • bulk-mechanical Well-specified repeating edits across many files. Renames, frontmatter patches, schema sweeps.
  • vault-scanner Hygiene passes across the vault. Broken links, frontmatter validation, orphaned notes.
  • web-researcher One leg of a multi-source investigation. Five running in parallel, each on a distinct slice.
  • Plan Architecture strategy. Implementation plans, trade-off analysis, step-by-step designs.

Consequence: parallelizable work runs parallel. Long searches don't burn your context budget. Fresh-head reviews actually come from a fresh head.

[ All five · traced once ]

One morning request. "What's my day?" Here's what fires, in order, without you naming any of it.

CLAUDE.md is already loaded. "Morning briefing" matches the session-start skill's trigger description.

The skill fires. It queries the Obsidian MCP for the Currently sections of each active project and the Gmail MCP for unread threads.

A rule from .claude/rules/journaling.md arms path-scoped: any file modification in this session owes a journal entry. A bulk-mechanical agent spawns in parallel to summarize three long Gmail threads without touching the main context.

Reply drafted. One new decision note logged. One journal entry started.

Structured knowledge is the byproduct. Every prompt leaves the vault a little more organized than it found it. A decision filed. A task opened. A journal landed. That output is what the next session reads. The system compounds while you work.

I didn't open anything to get those answers. I asked once and the system went and looked.
[ The loop ]

Every session leaves the structure a little more organized than it found it. A decision gets written down. A note gets linked. An old assumption gets corrected.

The next session starts from there. More to read, sharper answers, more worth writing down.

Six months in, both me and Claude have access to [N] cross-session decisions, [N] journal entries, [N] typed knowledge notes, [N] skills, all created in the course of working, none of it filed by hand.

The knowledge didn't decay. It compounded.

[ The close ]

You could build this. Everyone who sees it can see how it works.

But you'd spend a year figuring out what to keep. I already did the year. The audit is me handing you only the parts that matter.

  • 30-minute call. I look at how your work actually flows and where an AI that remembered everything would save you the most hours.
  • One concrete workflow we'd rebuild first. The specific thing to tackle, not a generic checklist.
  • Written recommendations you keep. Usable whether we work together after or not.
  • An honest read on whether you should DIY. If your situation's simple enough, I'll say so and point you at where to start.

You don't have to build this yourself. I've already built mine. I can build yours to fit your work.