# BLINDSPOTS.md — deep-dive meta-audit of goal-prompts - date: 2026-07-10 - method: five parallel audit tracks (UX-from-rendered-pixels, build architecture, catalog content, product strategy, adversarial cross-cutting sweep), synthesized. Not the output of a catalog brief; a one-off meta-audit of the whole product. - scope: every system and surface — site, briefs, build, distribution channels, business pages, and the repo's own decision/report corpus. ## 0 · The one-paragraph verdict The product is a beautifully machined answer to the wrong layer of the question. Its own dogfood reports already knew this: DEMAND.md locates willingness-to-pay in the execution layer and "~Zero" in the catalog layer; COMPETITIVE.md names the headline gap as "distribution and proof, not features"; docs/usage-metrics.md designs the usage instrument and states "Nothing in this document is live." The repo then shipped ~50 storefront optimizations while the three items that create an audience (npm publish, launch, measurement) stayed open and the pivot decision (ROADMAP R62) stayed permanently deferred. Every blind spot below is a variation on one failure mode: **verifying the verifiable proxy instead of the value** — lint the skeleton not the efficacy, guard the conductor text not the search semantics, optimize the funnel not the traffic, mark the finding fixed without wiring the fix. ## 1 · The unknown knowns — admitted in this repo's own documents, ignored by the product Each of these is written down, in this repo, by this project's own process — and the shipped state contradicts it. 1. **"The money is one layer over" — decision never taken.** DEMAND.md's closing verdict (execution wedge vs catalog) was filed as ROADMAP R62 "external (founder decision)" and never ruled on. ADR-9 answered a different question (Build-family products) without ruling on the wedge. 2. **"Distribution and proof, not features" — then a 50-item feature sprint.** COMPETITIVE.md §1 vs ROADMAP's execution log: R01–R52 shipped; R59 (npm publish), R60 (launch/Discussions), R61 (demo recording) — the only audience-creating items — remain open. metrics.json: 0 stars, 0 forks. 3. **The usage instrument is designed, zero-code, and off.** docs/usage-metrics.md ("Nothing in this document is live"), middleware.js.example ("INERT ON PURPOSE"), five separate reports flagging raw-fetch counting as the gating instrument (FUNNEL §4.3, REVENUE §4.4, RETENTION §4, PROOF §4, COMPETITIVE §4). CLAUDE.md sells raw/ fetch counts as "an honest usage metric"; nothing counts them. Consequence: the maintainer cannot answer "which briefs get run" for ANY channel — plugin/installer/MCP users generate zero observable events. 4. **ADR-9 parked monetization; /teams and /partners shipped anyway.** The append-only decision log — the asset the whole harness thesis depends on — is contradicted by build.py's teams_page/partners_page, selling against an audience REVENUE.md calls "unproven and unmeasured," with a partner CTA pointing at GitHub Discussions, which is not enabled. 5. **The flagship loop leaks on its own repo.** ACTIVATION.md §0: a finding was dispositioned **fixed** while "the fix shipped everywhere but where the finding pointed." RETENTION R12: gp-aid shipped undocumented against the prior report's explicit instruction. The report→Fixer→FIXLOG loop is the product's differentiation and it demonstrably drops findings. 6. **Briefs can produce confident, verifiable, worthless output — and nothing evals them.** ADR-12: ten design passes "produced changes users could not perceive… design edits shipped on plausibility alone." The fix (render the pixels) was applied to this repo's own workflow only — never generalized to the 146 briefs sold to users. FABLE_BUILD_QUEUE item 2 admits rubric evals are the missing layer; unstaffed. 7. **The harness has never been traversed.** HARNESS_PLAN §8's done-criteria (one product repo, first dollar, three HEALTH.md rows) don't exist; FABLE_BUILD_QUEUE item 1 admits "the first traversal is the eval of the harness itself." ADR-9's revenue theory rests on an untested path. ## 2 · The underlying logical flaws - **F1 — A conversion-shaped storefront with no traffic and no conversion signal.** The site is meticulously instrumented for on-page events while the product's actual conversion (a brief getting *run*) happens off-site, 10–20 minutes after the last click (FUNNEL.md), on channels that emit nothing. Optimizing a ratio whose numerator is invisible and whose denominator is ~0. - **F2 — The report is treated as the product; the user wants the change.** The read-only-report formula is the catalog's identity, and the repo has already quietly invented every escape hatch — 47 (fixer), 29 (monitor), 142 (generator), 144 (scored gate) — without promoting any to a first-class brief shape. 123/146 briefs use the literal header "Audit through N lenses"; 78 land on exactly 7. The formula is audibly mechanical. - **F3 — Static prompt text is the commodity layer.** The durable assets here are the report grammar, the linter, the diff/re-run rule, and the Studio→Fixer loop — all of which point at an execution product (scheduled runner, trend dashboards) that exists only as an example YAML file and one page (vitals) reachable mainly from the footer. - **F4 — Quality gates verify shape, never substance.** The linter checks the 4-phase skeleton, char cap, and ask-first gate; a hollow brief that pattern-matches passes clean. No seeded-defect evals, no cross-brief similarity check, no consumer-side "how to judge a report" guidance anywhere. - **F5 — "stdlib-only" drifted into "one language" dogma.** Node is already a first-class runtime (MCP server, test suites, syntax gates), yet search logic is triplicated across Python/Node/browser with a string-matching parity guard that misses the part that actually diverged. The honest constraint is "no runtime deps, no build-time npm install," not "one language." - **F6 — Strategy by accretion.** 7 distribution channels, 37 playbooks, 21 families, 146 briefs — each addition individually defensible (COMPETITIVE.md "table stakes"), collectively an unprioritized surface that multiplies maintenance and dilutes the story. The repo ships a *Subtract* family and has never applied it to itself at the product level. ## 3 · Architecture blind spots (ranked, with effort) 1. **Search semantics silently diverge across runtimes.** Site: substring + edit-distance fuzzy (js/catalog-core.js); MCP server: word-boundary regex, no fuzzy (mcp/server.cjs:157). The smoke-test parity guard checks conductor *text* only. Fix: add `js/` to npm files, make server.cjs require catalog-core, delete the duplicates and the guard. ~half a day; highest leverage in the repo. 2. **Detail-page inline JS is never syntax-checked.** scripts/check parses inline scripts in index/studio/vitals only; the 183 generated b/ and p/ pages carry Python-string JS with no gate. Fix: extend the vm.Script loop to one sample of each. ~10 lines. 3. **~17KB of SITE_CSS inlined into each of 183 committed pages** (~3MB duplication; one CSS tweak rewrites 183 tracked files, destroying diffs and blame). Fix: emit site.css and link it (tokens.css already proves the pattern). ~1 hour. 4. **Page chrome exists in 4+ hand-maintained copies with no guard** — the theme-toggle script alone has four literal copies (build.py, template.html, studio.html, vitals.html). Fix: shared chrome.css + js/theme.js, or generate studio/vitals heads from build.py. ~1 day. 5. **A growth ceiling encoded as a build failure.** The 16-stage conductor cap (build.py:205) × auto-built family conductors: Design is at 15 briefs; the 17th fails the whole build with no splitting mechanism. Fix: auto-chunk family conductors. ~half a day. 6. **The catalog ships twice in index.html** (injected JSON + server-rendered cards, 228KB), and the SW precaches bodies.json (420KB) + 5 fonts for every visitor on first paint. No pagination story at 300 briefs. ~1–2 days. 7. **Behavioral test void.** 1,320 lines of inline template.html JS get only a parse check; js/gp-detail.js (runs on all 183 detail pages) has no suite; vitals.html reimplements its own untested markdown-table parser instead of using the *tested* js/report-parser.js. One jsdom smoke of render() covers more product surface than everything else combined. 8. **build.py is a god-module** (8 responsibilities, 2,274 lines). Mechanical split into lint/render/assets. ~1 day, no behavior change. 9. **Ungated assets institutionalize silent staleness** — metrics.json and img/studio.png refresh out-of-band with no age check; committed zips bloat .git (26MB) monotonically. ## 4 · UX / navigation findings (from rendered pixels) **P0** 1. **Mobile nav amputates the product.** At ≤720px the nav drops How-it-works and *Catalog*; ≤480px drops the CTA; no hamburger. Studio/Vitals/Examples show **zero** nav links on mobile. The mobile landing page is 12,418px tall and the only route to the catalog is scroll. Fix: mobile disclosure menu; if pruning, keep Catalog + CTA, drop the secondaries. 2. **The proof surface dead-ends into raw markdown.** examples/ cards link to /reports/*.md — unstyled plain text, no nav, no way back. The single most persuasive artifact renders as a text dump. Fix: render reports to styled HTML (the CHANGELOG→changelog.html machinery already exists) or open them in the Studio. **P1** 3. **Search silently collapses recall.** Landing search runs on metadata only; bodies merge in only on *zero* results — "top 1 for 'security'" while the Trust chip says 11. Fix: idle-prefetch bodies.json and always search full text (it's already SW-precached), or upgrade when results < N. 4. **Half the site hangs off the footer.** Vitals, examples, quality, teams, partners, changelog have no top-nav or breadcrumb presence; b//p/ pages get breadcrumbs, tool pages don't. Promote Vitals next to Studio (they're a pair) and unify the breadcrumb pattern. 5. **The playbook storefront is an unexplained chip wall** — ~31 bare chips after the 6 featured cards, no descriptions or grouping. Group by intent or reuse the quick-view pattern. 6. **Two different sites' navs** — landing (uppercase mono, "Browse catalog") vs every generated subpage (sentence-case sans, "Get started"). Per DESIGN_DIRECTION rule 4, the subpages are off-direction. **P2** — chips/buttons below 44px touch size; first-viewport text-wrap glitches (hero callout, orphaned ▸ chevron); quick view opens an unframed full-viewport text wall with the close affordance scrolled away; keyboard nav (/, Cmd+K, j/k) fully wired but hinted nowhere; Studio/Vitals open empty instead of auto-loading the sample (the "no visualizations" complaint is mostly this); the example filename — the row's concrete payload — is what mobile truncation cuts. **Counter-evidence (complaints that don't reproduce):** progressive disclosure exists (the pick-a-question gate before the card wall); motion is present, tasteful, and reduced-motion-aware; copy feedback (toast + mark-it-run) is excellent; no horizontal scroll at 390px; dark mode is clean everywhere; the "ledger" direction is being executed *well* — the problems are navigation plumbing, search recall, and empty-by-default tools, all fixable inside the existing direction. The one direction-level friction is mono-metadata density at 12px on small screens, which needs a mobile relaxation, not a redesign. ## 5 · Cross-cutting orphans 1. **Fetch-and-execute trust chain.** Conductors tell agents to curl live URLs and "execute exactly as written," then 47/Studio-Fixer modifies code; the installer's checksums come from the same origin they'd defend against. No version pinning, no per-brief hashes, no threat model. (The MCP server's local input validation is clean.) 2. **Copied prompts are unversioned snapshots.** The installer and MCP stamp versions; the two most-promoted flows (copy button, raw/ URLs) don't — so a report can never be traced to the brief text that produced it, and the linted "lead with what changed" re-run rule can't distinguish brief drift from repo drift. Fix: one provenance line in raw bodies + report footers. 3. **No vulnerability-disclosure path** — no SECURITY.md anywhere, and vercel.json permanently redirects /SECURITY.md to a marketing audit report. build.py even reserves SECURITY.md as a protected community filename. 4. **gp-aid vs the privacy claim.** A persistent per-device id attached to every analytics event, plus logged search terms, alongside "no cookies, no cross-site tracking" — GDPR-wise a durable identifier is consent-scoped like a cookie. No privacy page, no opt-out, no rotation. 5. **Zero accessibility verification while selling an a11y brief** (86). Hand- authored a11y is genuinely good; nothing guards it — no axe/pa11y in scripts/check, and among 28 dogfood reports there is no A11Y.md. 6. **Consumer-side report judgment is nobody's job.** Nothing anywhere says "spot-check citations before ticking Fixer boxes"; hallucinated-but-well- formatted findings flow straight into code changes. 7. **Plugin staleness** — no version anywhere in plugin command files; no in-product "you're N versions behind." 8. **manifest.json drifted from the ledger palette** (#131417 vs #15120D), no maskable icons, no build gate — unlike the count-guarded OG cards. 9. **i18n has neither an implementation nor an ADR** saying "deliberately never" — the only major axis with no recorded decision. **Clean (checked):** stars badge threshold-gating, README count gates, CI drift guard (airtight), sitemap/robots, CHANGELOG discipline, Studio parser's honest degradation on malformed input. ## 6 · Content: catalog health, gaps, and shapes - **Overlap clusters to merge or signpost:** Design micro-lenses 54–59 + 132 (six audits citing the same CSS); Growth conversion path 09/75/80/109/110; dependency pair 07/85; findability quartet 14/16/52/145; testing quartet 02/100/101/102. The hub-and-spoke signposting that 118 does well ("for the defensive review, run 35") is missing exactly where it's needed; `related:` is sparsely used in these clusters. - **Real coverage gaps (ranked by demand):** mobile-native (iOS/Android — zero coverage), supply-chain security (lockfiles, provenance, CI secret exfil — distinct from 07/85's dependency *quality*), classic ML/data-science (the Agent family is LLM-only), pricing & packaging, background jobs & queues, alert quality/pager health, realtime & offline sync correctness, support ops. Honorables: experimentation rigor, cloud-bill/IaC cost, docs-site IA. - **Non-gaps (verified before claiming):** frontend perf, i18n, a11y, incident response, repo DX, DB perf, licensing, email, search, IaC, growth loops. - **Shapes to promote from exception to first-class:** **Monitor** (29 is the prototype; the linter already forces the diff grammar on every brief — vitals is its surface), **Generator** (Phase-2 scope gate like 47/142, deliverable is the artifact: the missing tests, evals, runbooks), **Scored gate** (144 generalized: bars before scoring, binary ruling, machine-parseable — what the Studio parser most wants). The 4,000-char cap is a non-issue (median 2,711). - **Choice architecture:** 146 briefs are navigable only through 37 playbooks, which are themselves too many to scan. Prune playbooks to ~15, designate a hub brief per large family, mandate `related:` inside overlap clusters. ## 7 · Are goal prompts the right implementation? **As content, yes. As the product layer, no — and the repo's own reports said so first.** The strongest case FOR the current form: zero backend, zero COGS, the only machine-enforced quality bar in the category, a parseable report grammar feeding a closed loop no rival owns, privacy as differentiation, and one markdown source compiling to every packaging that survives harness churn. The strongest case AGAINST: DEMAND.md — willingness-to-pay "High for the execution layer, ~Zero for the catalog layer"; "nobody is asking to buy a prompt catalog"; platform absorption underway (first-party /security-review, skills, workflows). Static prompt text is the commodity layer, and the paste model structurally blinds the product to its own value delivery. The 10x version is already half-built in four disconnected places: the GitHub-Action runner (one example file), report diffing over time (vitals + the universal dated re-run rule), brief evals (template/evals/run.py exists for downstream products, never pointed at the briefs themselves), and team rebasing (GOAL_PROMPTS_BASE). The composition — **a runner that executes versioned, eval-tested briefs on a schedule and renders repo-health trends from the diff grammar, with the free catalog as its acquisition engine** — is the product the evidence supports. ## 8 · Ranked action plan 1. **Take the R62 decision.** Rule on the execution wedge explicitly, in DECISIONS.md. Everything else sequences behind this. 2. **Flip on measurement + ship the three external unlocks** (raw-fetch counting per docs/usage-metrics.md, npm publish, launch/Discussions). Hours of work; currently gating every other prioritization. 3. **Fix the P0/P1 UX plumbing** (mobile nav, styled report pages, full-text search, footer-orphan surfaces, playbook chip wall, nav consistency) — all inside the existing design direction. 4. **Build brief-efficacy evals** (seeded-defect fixtures + expected-findings assertions; publish pass rates on /quality). Converts the quality claim from shape to truth; answers the ADR-12 failure class. 5. **Pay down the architecture quadruplication** (shared search module, linked site.css, shared chrome, detail-page JS gate, conductor-cap chunking). 6. **Commit to the diff loop as the retention spine** (generalize HEALTH.md/vitals to every report re-run). 7. **Prune: channels, playbooks, overlap clusters; enforce or supersede ADR-9** (teams/partners down to a contact line until adoption exists). 8. **Close the trust orphans** (SECURITY.md, provenance stamps on copied prompts, consumer-side report-judgment guidance, gp-aid privacy page).