# Agentic profile — Luca
lucadebort@pm.me · Bologna · freelance
Window: 2026-05 → 2026-06 · 41 sessions · 805 instructions · 4 products
Log consistency screen: 100/100 (screen, not proof)
Sources: claude-code (full capture) · 41 sessions read
Window and counts are lower bounds: logs rotate and old sessions are pruned.

A freelance designer and semiotician who also ships production code, building consumer and creative apps across web and native iOS, porting prototypes into design systems, and crafting presentation decks. Works as a daily driver, mostly co-thinking: detailed briefs to launch greenfield work, then tight iterative refinement. Returns to a handful of products repeatedly, depth over breadth.

## Domains
- **Web and native application development** · 1 products · 19 sessions — consumer and creative apps across web and iOS, plus a design-to-code sync toolchain; 111 commits
- **Design-systems and design-to-code engineering** · 1 products · 11 sessions — porting interactive prototypes into a maintained design system; localization of structured content
- **Presentation and talk design** · 1 products · 9 sessions — an HTML slide deck on AI-assisted qualitative analysis, with build tooling
- **Developer tooling and environment configuration** · 1 products · 2 sessions — auditing and repairing a local agentic setup

## Representative projects

### Consumer and creative application development across web and native iOS, plus a design-to-code synchronization toolchain.  ·  static-site · design-research _(p1 — projects)_
2026-05→2026-06 · 19 sessions · commits 111 · reverts 3 · churn med · checks passed
stack: Swift, Python
Built and iterated a portfolio of apps over 19 sessions and 111 commits: a native iOS dream-journaling app, a divination and reflection app, web-based data-visualization and utility tools, and a greenfield monorepo for keeping design and code in sync. Heavy focus on visual polish, animation timing, and App Store framing.
_why representative:_ The largest and longest-running effort, spanning native, web, and tooling work and the bulk of the candidate's commits.
artifact: Design-system component library (live)

### Design-systems engineering for client deliverables: converting interactive prototypes into a maintained design system, plus localization of structured content.  ·  feature-work · design-research _(p2 — wërk)_
2026-05→2026-06 · 11 sessions · commits 5 · reverts 0 · churn low · checks passed
stack: Python
Across 11 sessions, reverse-engineered an interactive prototype presentation into design-system foundations and components, working bidirectionally between code and design. Also handled localization of structured survey content, repeatedly auditing the model's output for changed strings and omitted fields, and shipped a small local-only time-tracking page with a cleaned-up design.
_why representative:_ Represents the professional, client-facing side of the practice, with tight verification, the deepest use of design-to-code tooling, and zero reverts.
artifact: Client prototype-to-design-system work (live)

### Presentation and talk design: an HTML-based conference deck and supporting build scripts on the theme of AI-assisted qualitative analysis.  ·  static-site _(p3 — talks)_
2026-05→2026-05 · 9 sessions · commits 24 · reverts 0 · churn low · checks passed
stack: Python
Designed and restructured a multi-slide HTML presentation over 9 sessions and 24 commits, iterating on narrative structure, visual layout, and staggered transitions. Built a Python pipeline to export to a standard slide format with animations, processed video assets, and deployed the result to a live site.
_why representative:_ A self-contained, shipped artifact that shows craft-level design iteration and a willingness to build supporting tooling.
artifact: HTML conference deck (talk-design sample)

### Developer tooling and local agentic-environment configuration.  ·  exploration _(p4 — unknown)_
2026-06→2026-06 · 2 sessions · commits 0 · reverts 0 · churn n/d
Two short exploratory sessions auditing the candidate's own agentic configuration, its installed skills and MCP servers, and troubleshooting a broken native-binary install. Maintenance work on the toolchain itself rather than a product.
_why representative:_ Included as the distinct exploration and tooling-maintenance type, separate from the work of building products.
artifact: — (none attached)

## Cognitive profile
tags: verification-heavy · risk-calibrated
Decomposes greenfield work up front through detailed implementation plans, then tracks execution with structured task lists and frequent clarifying exchanges before committing to a direction. Verification is a defining, rising habit: the candidate inspects model output against the source, catching silently changed strings, omitted fields, and prototype-to-screen divergence rather than accepting work at face value. Risk tolerance is calibrated by context, comfortable letting the model push freely on personal projects (some reverts) but tight and low-churn on client deliverables (zero reverts). Orchestration leans on a custom MCP toolchain and occasional sub-agent delegation, but the candidate stays in the loop, course-correcting in short cycles. Trust is selective: close direction on architecture and specs, lighter co-thinking through iterative design.

## How they work with the AI
25% directing · 75% co-thinking · co-thinking-leaning
Leans co-thinking, roughly three quarters of exchanges. Switches to directing mode to launch work: detailed implementation plans, thorough codebase audits, and full product briefs set the frame. Once building, the candidate co-thinks, steering through short, colloquial reactions, corrections, and continue prompts rather than fresh specifications.

## Practice intensity
- Active days: 30 / 45 (67%)
- Median sessions per active day: 2
- Median session depth: 43 tool calls
- Longest streak: 13 consecutive days
- Peak day: 1058 tool calls
Cadence: daily driver · focused sessions

A daily driver: active on 30 of 45 observed days, including a 13-day streak, typically two focused sessions a day. Sessions run a median of 43 tool calls and peaked above a thousand in a single day, so Claude sits at the core of the workflow.

## Work distribution
- Sessions per product: median 11 · mean 10.3
- Top 3 products: 95% of all sessions
- Multi-month products: 2 / 4 (50%)
Shape: deep focus

Deep focus rather than broad steering: just four products, with the top three holding 95% of sessions and two sustained across multiple months. The candidate returns to a small set of efforts repeatedly, building them out over time rather than sampling many briefly.

## Agentic literacy
Uses
- Sub-agent delegations: 8
- Task tracking events: 153
- Built-in slash commands invoked: 12
- Custom skills/commands: 6 distinct, 17 invocations
- Public MCP servers: 1 · 125 calls
- Custom MCP servers: 1 · 8 tools · 303 calls

Builds
- Skills authored: 1
- Commands authored: 0
- Agents authored: 0
- Hooks edited: 2
- Project memory files (CLAUDE.md): 5

Designs
- Plans-first (ExitPlanMode): 2
- Subtask tracking (TodoWrite): 186
- Clarifies (AskUserQuestion): 57

Operates a mature agentic stack: has authored a custom skill, edited hooks, and maintains project-level memory files. Integrates a self-built multi-tool MCP server driven through hundreds of calls alongside a public one, and reaches for custom commands, sub-agent delegation, and structured task tracking as standard tooling rather than novelties.

## Trajectory
Over the window the practice shifted from exploration toward confident execution. Prompts grew more specified (median 11 to 15 words) and delegation roughly doubled, while the ratio of research to actual changes fell (0.75 to 0.6), meaning less circling and more building. Verification rose in step, so faster execution did not come at the cost of scrutiny. Throughout, the work stays overwhelmingly design and UI centered, with developer tooling only a minor strand.

Behavioral shifts (early → late half, split at 2026-05):
- median prompt words: 11 → 15 ↑
- delegation rate: 1% → 1% ↑
- research:mutation: 0.75 → 0.60 ↓
- verification rate: 22% → 26% ↑

Topics explored, by quarter:
- 2026-Q2: design & UI (20), dev tooling (1)

New vocabulary adopted: tailwind, mapping, layers, package, detection, interpretation, inference, enrichment, audits, eiid

## Stack adopted
Swift, Python
_Stack as currently declared in the repo: from package.json dependencies (npm manifests only) and the files touched. Dependencies are unavailable if the project folder no longer exists, and services used only over REST (no package) are not auto-detected._