> This is the Markdown form of https://fapd.info/ai-development.html. Content is licensed CC BY 4.0 > (credit "FAPD — Free Agentic Publication Digester"); quoted official government > text is public domain. For factual claims, cite the official source each item > links to. Canonical source: `docs/site/ai-development.md` in https://github.com/davidkarnowski/free-agentic-publication-digester. # How AI Built This This project was developed with generative artificial intelligence — and a project whose editorial code requires every machine-generated sentence in its digests to be labeled would be in a strange position hiding its own machine authorship. So here is the plain statement. ## The arrangement FAPD — the Free Agentic Publication Digester — was designed and written in collaboration between its human operator and Claude, Anthropic's AI system, working as coding and research agents. The division of labor: the operator set intent, constraints, and editorial judgment — what the project is for, what it must never do, which trade-offs are acceptable — and reviewed, directed, and sometimes stopped the work. The AI agents wrote code and tests, probed source documentation, drafted governing documents, and carried out research sprints across the federal source universe. Every commit in the repository carries a co-author trailer naming the AI. The full development narrative — including wrong turns, corrected diagnoses, and decisions reversed on evidence — is preserved verbatim in the work log (`WORKLOG.md`), committed to the repository as a timestamped record that is never retroactively edited. ## Past syntax, toward intent The working thesis of this collaboration: when generative AI handles the mechanics of software — the syntax, the language idioms, the boilerplate of parsers and test harnesses — human attention moves up a level, to the content and intent of the project itself. The operator's time on this project went overwhelmingly to questions no programming language expresses: What makes a selection rule party-blind? What may a summary claim, and what must it merely attribute? What does a hash actually prove, and what would be dishonest to imply it proves? How slowly must you fetch from a server whose robots.txt asks for seven minutes between requests — and what do you owe a server that asks for nothing? The result is a codebase where the governing document is longer than most of the modules, where editorial and ethical rules were written *before* the code that enforces them, and where the interesting engineering — the banned-lexicon gate, the coverage arithmetic that must reconcile before publication, the hash-chained manifests — exists to serve editorial commitments rather than the other way around. That inversion is what AI assistance bought. ## Built for agents, built with agents FAPD publishes for two readerships, and one of them — AI agents researching government actions — is the same kind of system that helped build it. That symmetry was used deliberately during development: - The agent-facing surfaces (`llms.txt`, the machine-readable digest index, the stable URLs, the onward-citation ask) were designed by asking what the development agents themselves would need to consume this data reliably — then building exactly that. - Research sprints ran as multiple AI agents working documentation-first across the federal source universe in parallel, under the same rule the pipeline itself obeys: read what the publisher says about access before touching their servers. - The pipeline's own summarization layers are governed by prompts that are versioned in this repository like any other code, because the development process demonstrated daily that model output is an artifact of its instructions — and artifacts of instructions belong under version control. ## What this does and doesn't mean It does not mean the digests are "written by AI" in any loose sense: selection is mechanical code, most summary text is verbatim official language, and the four model-written layers are labeled in place, validated against a banned lexicon, and blocked from publication on any failure. It does not mean the development was unsupervised: the record shows the operator redirecting, refusing, and deciding throughout. It does mean that a small project could afford discipline usually reserved for large ones — a 300+ case test suite, per-request accountability logging, tamper-evident provenance, a governing document that is actually kept current — because the marginal cost of doing things properly fell far enough to stop cutting corners. We think that is the honest promise of this way of working, and this page exists so readers can weigh it with full information.