The values I trust are the ones I can point at: to a paragraph in a case study, a line of code, a thing I cut and the reason I cut it; the rest are vibes. Five things I keep coming back to, named only after each had shown up in the work three or four times: not a manifesto, just a working list. Each links to where it earns its place, and to the reading that grounds it. Below the values, four reusable methods I keep reaching for: not new philosophy, just the patterns I use when the stance has to become an action.
01
Name the thing first.
Pick the name and the metaphor before you pick the framework. Whatever you call the thing decides what gets built next.
Spec before component, schema before viz, periodic table before grid. Sprint protection before sprint planning (SPF). The metaphor decides what's load-bearing. A name that doesn't constrain you is wallpaper.
Reading: Austin & Searle on speech act theory · Christopher Alexander, The Timeless Way of Building
↻ This card was named first. The body came after the title; the title is what made the body writable.
02
Let the file enforce the rule.
AI tooling makes building cheap. The actual work is keeping past decisions decided, in one file the tools have to read on every change.
A rule that lives in a file the system reads on every change beats a style guide nobody reads. The looper, cut on Omunikorudo. The third sprint column, cut on SPF. The per-task estimator, also cut on SPF. They all stayed cut because the file got consulted at every commit, not at next quarter's design review.
Reading: Hofstadter, I Am a Strange Loop · Luhmann's Zettelkasten as ancestor
↻ This card lives in src/pages/values.astro. Edit the file, the page changes. No second style guide to drift away from.
03
Refuse the easy lie.
The fast version of anything is almost always a lie: a fabricated name, a smoothed corner, a placeholder you forgot to remove. The honest version costs more. That cost is the receipt.
Every sentence on flowerpostcards is real, pulled from a 452-entry corpus, never paraphrased. The Lab Notebook on the Periodic Table is permanent; reversals get noted in place, not overwritten. Pocketbook captions are hand-edited after the model writes them. And this site's copy contains no fabricated names. The omission is the rule.
Reading: the site's own no-fabrication rule, written into the obsessions
↻ This card has one reading entry, not two: I declined to invent an author I'm not sure I've read. The shorter list is the receipt.
04
Show the working.
Keep the lab notebook, the dated change log, the rejected drafts, the populated demo: visible. People trust polished work whose mistakes are still readable underneath.
The most-loved page on the Periodic Table is the Lab Notebook: a never-edited record of every reversal, kept the way kids of the '90s web kept view-source on. The most convincing thing about Pocketbook is the diff: the captions show their human edits in version control, not just the polished result.
Reading: Annie Dillard, The Writing Life · the kid-of-the-90s-web habit of leaving view-source on
↻ Every claim above links to the project where it actually shows up. View-source is on, and the card is the receipt.
05
Build it to outlast.
Considered work, not volume: a build stamp on every page so the site admits its own date. It's here to last, not to harvest.
No tracker, no analytics, no cookie banner, no signup gate, no harvest. Build stamps on every page so it ages honestly, instead of pretending to be present-tense forever.
Reading: Octavia Butler, Earthseed ("the only lasting truth is Change") · Lewis Hyde, The Gift
↻ The colophon at the bottom of this page dates this card. By the time you finish reading it, it isn't anymore.
From stances to patterns
Methods
The five above are stances. The four below are how I act on them: reusable patterns the work keeps returning to. Each links to a fuller write-up with a proof box, the trade-offs, and the projects it runs on.
M01·Method·Hireable · Forward-thinking
AI-era Product Development Life Cycle
Idea to spec to build, when AI compresses the build half
A practical operating pattern for moving from idea to spec to build when AI speeds up production but raises the cost of unclear judgment. Faster building makes weak framing more expensive.
The prompt is not the trick; the discourse boundary is the product
Designing prompts as language systems with boundaries, evidence rules, exclusions, and intended speech acts. Language behavior needs product rules, not vibes alone.
Separating canonical memory from working notes, experiments, and archived loops, so a system can stay useful over time. AI-assisted work needs grounded memory, not an infinite junk drawer.
Model the constraints first; let the model do the explaining
AI is better when it explains a trustworthy structure than when it invents the structure in real time. Constraints should be modeled before they are explained, especially in regulated work.
