diff --git a/docs/superpowers/specs/2026-06-01-blog-search-design.md b/docs/superpowers/specs/2026-06-01-blog-search-design.md
new file mode 100644
index 0000000..db45cec
--- /dev/null
+++ b/docs/superpowers/specs/2026-06-01-blog-search-design.md
@@ -0,0 +1,124 @@
+# Blog search — design
+
+**Date:** 2026-06-01
+**Status:** Approved (pending user spec review)
+**Scope:** Add a light, in-page search to `/blog` that filters the existing archive by keyword.
+
+## Goal
+
+Let a reader find a blog post on the `/blog` index by typing a keyword or short phrase. Match against the post's title, tag, standfirst/description, and body text. Filter the existing year-grouped archive in place.
+
+## Non-goals
+
+- Ranking, fuzzy match, excerpt highlighting (upgrade path: Pagefind).
+- URL persistence of the query.
+- Sitewide / top-bar search.
+- Tag-click-to-filter (separate feature).
+
+## Approach
+
+A tiny client-side filter implemented as **data attributes on each row** (no JSON index, no fetch, no React island).
+
+Each rendered post element carries a `data-search` attribute containing a lowercased, whitespace-collapsed concatenation of its searchable fields. A small inline `<script>` on the blog index reads the input, tokenizes the query, and toggles a `.is-hidden` class on every element whose `data-search` doesn't contain all tokens.
+
+Rejected alternatives:
+
+- **Inline JSON index + slug→row mapping**: viable but adds a parsing/lookup layer for no real benefit at this scale.
+- **Fetched `/blog/search-index.json`**: only pays off past ~50 posts; reconsider then.
+- **Pagefind**: full-text engine with ranking; ~70KB+ WASM. Overkill for current archive size. Document as the upgrade path.
+
+## Components
+
+### New: `src/components/molecules/BlogSearch.astro`
+
+Renders a single `<input type="search">` with a small inline `<script>` (vanilla JS).
+
+- `aria-label="Search posts"`, placeholder `Search the journal…`.
+- Right of the input on focus: a mono `Esc to clear` hint.
+- Width: caps at ~28ch desktop, full-width mobile.
+- Styling: `font-mono` 13px, hairline bottom border using `--hairline-on-dark` (light/dark variants like existing `.tag` / `.date`).
+
+The inline script (~30 lines) does:
+
+1. On `input`, normalize query: `value.toLowerCase().trim().split(/\s+/).filter(Boolean)`.
+2. For each `[data-search]`: visible iff every token is a substring of its `data-search` value. Toggle `.is-hidden`.
+3. The featured block (marked `data-featured`) is additionally `.is-hidden` whenever the query is non-empty, independent of match.
+4. For each `[data-year-head]`: count visible siblings until the next year head; hide the head if zero.
+5. Toggle a no-results message (`<p data-empty>`) on/off based on whether any row is visible.
+6. Empty query: clear `.is-hidden` from everything (including featured); hide the no-results message.
+
+Keyboard:
+
+- `/` anywhere on the page focuses the input (skip if a text input is already focused; check `document.activeElement`).
+- `Esc` while focused: clear value, blur, restore full archive.
+- `Enter`: no-op (filter is live).
+
+No debouncing — per-keystroke iteration over a few hundred elements is sub-millisecond.
+
+### Edited: `src/pages/blog/index.astro`
+
+- Build `searchText` per post in the frontmatter: `[title, tag, standfirst, description, post.body].filter(Boolean).join(' ')`, then lowercase, strip markdown punctuation (`#*_>\`[]()`), and collapse whitespace.
+- Mount `<BlogSearch />` between `.masthead` and `.feature`.
+- Add `data-featured data-search={featuredSearchText}` to the featured `<article>`.
+- Pass `searchText` to each `PostRow`.
+- Render `<p data-empty hidden>No entries match "<span data-empty-query></span>".</p>` after `.archive`, styled like `.count` (mono 11px, `--fg-on-dark-3`), `aria-live="polite"`.
+
+### Edited: `src/components/molecules/PostRow.astro`
+
+- Accept a new prop `searchText: string`.
+- Emit `data-search={searchText}` on the row's root element.
+
+### Edited: `src/components/molecules/YearHead.astro`
+
+- Emit `data-year-head` on the root element so the script can find heads.
+
+## Data flow
+
+Build-time: each post's frontmatter and body are concatenated, normalized, and embedded as a `data-search` attribute string on the rendered row. No runtime parsing.
+
+Runtime: input event → tokenize → DOM walk → toggle classes. All synchronous. No fetch, no JSON, no async.
+
+## CSS
+
+Add to `src/pages/blog/index.astro`'s scoped styles:
+
+```css
+.is-hidden { display: none; }
+```
+
+Empty-state `<p>` reuses the `.count` style (mono 11px, `--fg-on-dark-3`).
+
+## Accessibility
+
+- Input has `aria-label="Search posts"`.
+- Empty-state message has `aria-live="polite"` so result-count changes are announced.
+- `.is-hidden` uses `display: none` so hidden content is skipped by assistive tech.
+- `/` shortcut respects existing focus — won't trap users typing elsewhere.
+
+## Edge cases
+
+| Case | Behavior |
+|---|---|
+| Markdown syntax in body | Stripped to plain text before joining; code fence contents kept (useful for searching function names). |
+| Whitespace-only query | Treated as empty: full archive restored. |
+| Tokens with regex-special chars | Safe — only `String.includes()` is used. |
+| Query matches nothing | No-results message shown; all rows + year heads hidden. |
+| Single-post archive | Featured shows with empty query; non-empty query hides featured and the lone archive entry. Acceptable. |
+| JavaScript disabled | Input renders but does nothing; full archive visible. No regression. |
+
+## Theme parity
+
+Input and empty-state colors follow the existing `body.dark` / `body:not(.dark)` pattern used by `.tag`, `.date`, `.count` in the current `index.astro`. No additions to `colors_and_type.css`.
+
+## Testing
+
+No test harness exists in the repo. Verification is manual via `npm run dev`:
+
+- Golden path: typing narrows results; year heads collapse when their group empties; clearing restores the full archive.
+- Edge cases listed above.
+- `/` keyboard shortcut, `Esc` clearing.
+- Light/dark theme parity.
+
+## Upgrade path
+
+If the archive grows past ~50 posts, or readers ask for fuzzy match / ranking / highlighted excerpts, swap this implementation for Pagefind. The `BlogSearch` component is the only surface that needs to change; the rest of the page is unaffected.