Inline content editor

Changing a headline or fixing a typo meant hunting through TypeScript, editing the right string literal without breaking the build, then redeploying. Small edits carried real friction — and real risk of shipping a mistake.

Rewriting a sentence is a different job from editing one, so the editor grew a second affordance for it. Describe the change in plain language — "make it more analytical", or one of four presets — and the rewrite comes back as a preview. Accepting it doesn't take a new path: the text is spliced into the same draft a manual edit would produce and saved through the same endpoint, so it lands as an ordinary reviewable diff and ⌘Z undoes it like any other change. Nothing is ever written without being shown first.

The first working version invented things. Asked to make "Defined outreach strategy and built the content stack" more specific, it returned templates, assets, and sequences — none of which were in the original. On portfolio copy that isn't a style problem but a false claim about my own work, and it reads plausibly enough to accept without noticing. The cause was structural: asking for specificity where there is none leaves nothing to do but invent. The fix was giving the model explicit permission to barely change anything — that returning something close to the original is a correct answer, and padding it with plausible detail is not. It all runs locally through the Claude CLI, so no content leaves the machine and there's no API key to manage.

Paragraphs, headings, and leads — plus single-field entries like page and section titles — no longer swap out for a boxed form. Click straight into the text and it becomes the input, styled to match, with a dashed underline marking edit mode; saving happens on ⌘Enter or by clicking away.

A case study serves two people at once: the recruiter with ninety seconds and the one who wants the research. "Tailor the read" gives a project three depths — quick, a few minutes, full — and each block declares the shortest read it belongs to. Deciding that block by block in a JSON file is guesswork, so the choice moved onto the page. In development every block grows an eye button in the margin, and whatever the current read drops is dimmed in place rather than removed — so you're reading the short version and the cut at the same time. One click hides a block from the selected read or brings it back, writing the tag to the content file through the same endpoint, guards, and reviewable diff as a text edit. Readers get plain CSS: no re-render, no second copy of the page, and nothing that hides a screen, a feature, or a section — the shorter reads trim the supporting detail around them.

Every affordance that makes the page editable also makes it a poor sample of itself: dimmed blocks, dashed add buttons, hover toolbars, a control in every margin. Preview turns all of it off from the toolbar — chrome hidden, click-to-edit disabled, highlight rings dropped, and the read depths switched from dimming back to genuinely hiding. What's left is what a visitor gets, on the same page, without a build or a deploy. It's a toggle rather than a mode you have to leave and come back from, so checking a change costs one click in each direction.