AI isn't my tool.
It's my partner.

AI approach

Three sources in a declared hierarchy, aggregated by a generator written with Claude Code: Cloudflare as the base truth because it needs no cookie banner (AI crawlers show up in the edge logs, not in Analytics), Search Console for demand, GA4 only for what people do, not how many of them there are. The system is under orders to say in one line when nothing happened. Two caveats, both mine: the quarterly spot-check is still limping, 4 of the 12 planned prompts, a single repetition; and the automatic Monday run delivered once out of four, the other three I launched the report myself.

Outcome

In the week-36 report retrievals are 232, but 74 are Claude-User and I most likely caused them myself while checking. The rest is almost all OpenAI, 61 from OAI-SearchBot and 41 from ChatGPT-User; the new voice is Perplexity, from 0 to 36. The week after, back to zero. And it does not cite me: with my name only ChatGPT does, without it neither of them.

AI approach

It lives as a Claude Code skill, and you don't write a system like this: you beat it into shape by using it. On test day real defects surfaced, found by looking at the output rather than rereading the spec. Two examples. The Filo, the signature element, is an SVG on the web and the browser does the anti-aliasing; in the PNG and the PDF it was a polyline, sampled at 26 points per segment. It had always been there; you could only see it outside the browser. And the typeface's optical factor, 0.78 in a document, measured on the actual font files came out at 0.86: today the system declares two, and says why, 0.86 on the web where parity is measured, 0.78 in print where it is set by eye.

Outcome

A value written in a document gets measured on the object before you use it. On 20 August the skill went from 2.4 to 2.9 in one day: the versions that matter were born from looking at the output, not from rereading the spec.

AI approach

My procedures live in a folder on my computer, and a small program, a hook, proposes them on its own when they apply: it reads the question and suggests the right skill, silent on everything else. It exists because a skill you have to remember to invoke is a dead skill: better one suggestion too many than silence on the day it mattered. On 14 September 2026 at 15:31 I ran my auditor over my 17: twelve operational, one with a caveat, four stopped. Fixed the same evening, and at 22:11 the count was 17 out of 17.

Outcome

My own check stopped four of my skills, design system included: the sections were there, but under headings it was not looking for. Fixed that evening, and the check fixed too. A rule is worth as much as its check, and the check is worth as much as the question it asks.

AI approach

A second, adversarial pass before delivery: twenty-one agents with a mandate to falsify the conclusions. Of the twelve load-bearing claims, five held, three were corrected, three fell, one was left without a primary source. Among the ones that fell was an arithmetic error of mine.

Outcome

The most expensive problem was written in plain sight in the terms of sale anyone could read. Before paying for due diligence, read the pages your supplier already publishes.

Real case, sensitive details omitted.

AI approach

Bespoke skills and hooks in Claude Code. What stays human forever is a list, not a promise: money, public exposure, people and deals, legal and tax, opening and closing projects. It cannot promote itself, and it can switch itself off but never back on. The bans are files, not sentences: tools denied by name, and the push rule covered by a test suite that tries to break it.

Outcome

Promoted to the second level at the first review, in September: 43 briefs out of 43, zero unrecoverable errors. The alarm that matters does not fire on one of its mistakes: it fires if for thirty days I merge more than 90% of its PRs within an hour without a comment. At that point the problem is me.

AI approach

Claude Code headless, running locally, at 07:00 on weekdays. The first autonomous run took seven minutes: it reads the git logs of every repo and the recent Claude sessions, infers what moved, updates the status files, opens the PR. Every inferred line stays marked as inferred until my merge confirms it; questions only about the gaps that live in my head or in real life. At the end it declares what it could not see instead of writing “nothing new”: so far, in every run, the artifacts, which never open in headless mode.

Outcome

I don't count how many PRs I merge, but how many I merge in under an hour without a comment: above 90% for thirty days, the problem is me.

Real case, sensitive details omitted.

AI approach

I built it with Claude Code, then put three layers of verification on top: in June the parity between engine and spreadsheet and an internal three-agent audit, and that same month a second audit that sent it back to me technically sound and economically stretched; in July a cross-review by a model from another vendor, with a mandate to tear it apart. Eight findings: two mechanical bugs, reproduced numerically before touching the model, and six plan choices. Today the guard is a sheet of 46 checks that declares itself red when it is red.

Outcome

At the end of June the monthly engine's parity was green across 52,560 cells, and it could not see the payout bug: taxes all fell in December, December profit went below zero, and dividends stayed at zero for ten years in every scenario. Engine and spreadsheet were wrong in exactly the same way: the logic gets audited separately.

Real case, sensitive details omitted.

AI approach

Generated from code, not written by hand: a build parses the vademecum files and assembles the app on its own. If the counts don't add up (80 KPIs, 366 notes phase by phase, 13 courses, 8 bias-watches, one per phase) the build fails on purpose: tested, one note fewer and it stops. Self-contained, no external dependencies.

Outcome

Decisions go through a checklist before they go through my gut, and a repeated check on the same decision shows the delta against the previous one. It is not public: it holds the actual projects I use it on, eight names that come back a hundred and thirty-two times across a hundred and twelve lines.

Real case, sensitive details omitted.

AI approach

Astro and Cloudflare. Writing goes through separate skills for voice and for form, The drafts closed in September went through two judges, facts and voice, with a score and a list of fixes to apply before publication. The privacy gate runs on the text and removes things: suppliers, server addresses, names.

Outcome

The article on the AI bubble, in its first draft, opened with a wrong sentence that three paragraphs rested on. I went back to the source before publishing, and it says so in the article. Speed is not the point: the check is.

AI approach

Four deep-research runs with Claude, in sequence, the last one re-checking the inputs of the first three: that is where the rent figure I had typed in myself, and it had taken at face value, came apart. Along the way it told me with data that a 10% net yield does not exist at any reasonable price and refused the appreciation I asked for; I caught it declaring the tenant absent, reading the PDF line by line as I would with a junior analyst. Break-even price and rent solved a second time by iteration, independently: gaps below 0.1%. The third run, through the public registers, found the seller in insolvency proceedings.

Outcome

A €42,000 rent leaves €7,668 in cash, 18%. Yield is not a property of the building: it is a function of the price you pay.

AI approach

A very long piece of work, in a chat with Claude: direct counts on the historical series instead of estimates, and the simulation re-run across five doses and in two variants, including the one that weakens my own conclusion. Backtesting I could never have done without AI.

Outcome

It is not the strategy that is dangerous: it is the dose. At full leverage 98.2% of paths are wiped out within ten years; the exact same trade at a quarter of the capital has practically zero ruin.

AI approach

Cowork with a broker connector and the state kept in a Project. The principle «the broker before the file» was born from a real incident, with the state file and the broker telling two different stories. The tool can prepare an order, not send it: the instruction is born inert and becomes an order only when I press the button inside the broker's platform.

Outcome

Daily monitoring without watching the market. No order leaves without my hand, and that is not a rule I gave myself: it is how the tool is built. Violations go into a register I reread with my AI in February.

AI approach

FastAPI backend and frontend written with Claude Code; the enrichment is done by Claude from the command line, with no dedicated API keys. The instructive part: the headless call inherited the interactive configuration, about 68 thousand tokens and one wasted turn, and fifteen positions out of forty came back failed; with four isolation flags it went from 72 seconds over three turns to 39 in one.

Outcome

A single view in a few minutes, once a month: it is the starting point of the review, not a dashboard to stare at every day.

Real case, sensitive details omitted.

AI approach

Prototyped with Claude Code, scoring in deterministic Python. Then the part that matters: a reviewer who had not written the code found five bugs, two of them critical, and every validation number was recomputed from scratch: the index subtracted, the historical S&P 500 constituents rebuilt (651 tickers, not just the survivors), on thresholds taken from the literature and not tuned on numbers already seen. The first backtest said 79.5% winners. Net of the index, alpha was negative, and the stocks it skipped returned more than the ones it picked. Survivor bias alone inflated returns by 11.1 points. The only signal that survived, quality, is worth +1.6% of alpha across 3,677 cases: you buy that with a factor ETF, not with a system.

Outcome

I enjoyed it, and I switched it off anyway. The baseline for a retail scanner isn't zero: it's buying the index, and net of costs and time it didn't beat it.

Real case, sensitive details omitted.

AI approach

Claude Code wrote the pipeline; at runtime there is no model, only deterministic Python and a template. A guard fails the build if an email address shows up in the aggregated file. A register of twenty assumptions in three states, confirmed, estimate, to be validated, each with its own source, and the register with the date of its last revision. The run is by hand, once a month.

Outcome

I automated the production of the document, not the decision: the board reads faster and decides with the same latency as before. I optimised the wrong bottleneck.

Real case, sensitive details omitted.

AI approach

The council of models rejected the first version of the method before it ran, 45 issues, and part of them I pushed back on. Then the scoring in deterministic Python and two independent runs, with the result of the other one declared upfront as a bias to manage. From an empty workspace to the final ranking in two hours and thirteen minutes.

Outcome

From gut to criteria: why that sector, and not another. And the framework failed its own threshold the same day: 48 out of 69 passed, far too undiscriminating.

Real case, sensitive details omitted.

AI approach

Course by course with Claude Code: the AI proposes the filter, I decide what stays, and what I cut stays on the record, so the cut is auditable. Final QA with 193 agents across the maps: 159 findings, 133 confirmed, 26 rejected. Sixteen were real defects and I fixed them; the rest were editorial cuts of mine, not errors.

Outcome

Not the diploma: eight questions, one per phase, and the shoals I have already run aground on.

AI approach

Claude with persistent instructions and the course materials uploaded, one project per exam: it cites file and page, separates "from your materials" from "general knowledge", and when I get something wrong it asks me a question instead of handing me the answer. Every now and then I ask it to audit itself against the slides: that is how I caught an incomplete formula, a wrong answer key and a mock exam with the correct answers almost all on the same letter. The pen stays for what has to be remembered: the micro and macro graphs I learned only by redrawing them by hand.

Outcome

At an advanced stage of preparation my mistakes were not about knowledge but about execution and time: you cure them by simulating, not by rereading. The method can be copied, the pace cannot.

AI approach

Claude Code to rewrite the pipeline and for the methodological challenge; the numbers in Python with golden tests, never from the model. Variable selection redone inside every single test, not once over the whole dataset, and kill criteria written before looking at the results; then a second, pre-registered test on 1,567 companies from another sector, a much larger one, over forty times the thesis sample.

Outcome

Balanced accuracy 50.0%, permutation p=0.498; on the large sample AUC 0.537, the same ceiling as a plain logistic regression: it isn't the method, it's the target. FAIL accepted: the protocol survives, the model does not. I had written the caveat about that number myself, in my thesis, and taking it seriously is another matter.

Real case, sensitive details omitted.

AI approach

Claude did the product and the delivery, not the storytelling: 18 recipes written and validated with a single-source refactor that closed a structural bug, 15 consistent generated images, a mobile audit, and a script with body, voice, slide and pause cues tuned for a B2-C1 speaker, with seven delivery rules and the table of words every Italian trips on. The hard part was not the site: it was taking an English pitch on stage as a non-native speaker.

Outcome

Exam passed with explicit compliments from the lecturer and the class. The same anti-slop discipline as this site, measured by an outside judgment.

AI approach

Three layers: a model that reads and classifies the documents, deterministic Python that does layout and totals, the PDF export. Confirmations are mandatory by design: the system asks where inferring would mean inventing.

Outcome

The value is not speed: in two cycles out of five the second pass cut the total by about a fifth: expenses already in the company's books that had ended up in the personal claim. The automation made the error visible, and every first-pass error became a written rule.

Real case, sensitive details omitted.

AI approach

Three mockups I had Claude Code make, rejected as a block: they all answered 'what's waiting for you', when the real question was 'where does each one stand'. The fourth passed on the first try; the final order, by last interaction, came from use. Forty minutes estimated before starting, zero euros.

Outcome

Building it costs almost nothing, keeping it costs attention: that is how one of the two predecessors died. That is why the death date lives in the app's counter: fewer than 7 opens by 22 September, and it goes off. It is not public: the real fronts are in there.

Real case, sensitive details omitted.