The problem is not the tool. The problem is the prompt architecture. When you feed a language model your project plan and ask for a full application, it regresses to the mean of every grant it has ever seen. Safe verbs. Generic outcomes. Zero risk. That is the opposite of what a strong application needs.

What works instead is a split-pipeline approach. Phase one is human-only: define the strategic narrative, the theory of change, the one sentence that makes a funder lean forward. Phase two is AI-accelerated: turn that narrative into the correct structural format, generate the budget narrative, draft the evaluation framework, and produce the compliance annexes.

At Nuoret Kotkat I built this pipeline for STEA and municipal applications. The LLM handled the repetitive scaffolding — formatting, cross-referencing objectives to budget lines, generating indicator tables — while I owned the narrative spine. The result was faster drafting and stronger applications, because the human energy went into the argument, not the word count.

The specific stack: Claude for long-context drafting, a custom prompt library with examples of funded vs. rejected applications, and a strict review checkpoint where no paragraph ships without a human coherence check. Speed without shortcut.

The first lesson: in a crisis, transparency is cheaper than secrecy. Every hidden decision becomes a rumour, and every rumour costs trust you cannot afford to lose. The team opened the books — literally. Tighter financial oversight, published reporting to the membership, plain explanations of where the money had gone and where it was going. The shock was real, but the alternative was worse: a federation that died because its leadership was too embarrassed to tell the truth.

The second lesson: keep what runs running. When an organisation has gone dormant the temptation is to redesign everything first. That is a trap. The work was sustaining what existed — keeping member relationships warm across four continents while the structures around them were being repaired. A 100-year network that lives is worth more than a perfect blueprint that does not.

The third lesson: cross-border diplomacy is operational, not ceremonial. Dialogue with organisations in conflict-affected areas through the Willy Brandt Centre Jerusalem meant showing up consistently with partners who had real reasons to distrust each other. The only currency that works there is consistency — showing up, following through, never promising what you cannot deliver.

I stepped out of the term at 23 with the federation stable, governance documentation in place, and member coordination systems carrying 500+ contacts forward. I would not recommend the experience, but I would recommend the lessons.

The insight behind Vantnod was to pull jurisdiction out of the core domain entirely. Instead of a Finnish accounting app that happens to work elsewhere, we built jurisdiction-neutral primitives: transactions, accounts, reconciliations, reporting templates. Each primitive is pure — it knows nothing about Finnish ALV or Estonian KM. Then we layer jurisdiction-specific adapters on top: a Finnish adapter that maps the neutral transaction to ALV codes, an Estonian adapter that maps to KM, a German one for USt.

This is not a new idea in computer science. It is the exact same principle that lets PostgreSQL run on any operating system: abstract the storage engine, plug in the OS-specific file handler. But in fintech, almost nobody does it, because the incumbent players grew up inside one jurisdiction and never had a reason to abstract.

The business consequence is speed. When we launch in Estonia, we do not fork the codebase. We write an Estonian adapter — roughly 3,000 lines of mapping and validation — and the core product stays identical. The same UI, the same LLM pipeline, the same bank reconciliation engine. Only the tax output changes. That is how you move at EU speed: one product, many jurisdictions, zero forks.