@clize/clize 0.35.2 → 0.35.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@clize/clize",
-  "version": "0.35.2",
+  "version": "0.35.3",
   "mcpName": "ai.clize/clize",
   "description": "The real-world capability layer for AI agents — domains, email, deploy, media, SEO/GEO. CLI + MCP.",
   "keywords": [

package/skills/clize-seo/SKILL.md

@@ -239,11 +239,13 @@ will — it cannot measure those. That half of the state exists only here.
 - **Every ledger file counts, and a plan is a queue.** A site may carry more than one ledger
   (tabledi: `worksheet.md` for en-US and `worksheet-global.md` for the other markets); rows in
   all of them are the site's rows, and a round that counts only one file will report "nothing
-  to build" beside 27 verified cells. A batch plan written under an earlier count rule (`≤5,
+  to build" beside the cells the other file already verified. A batch plan written under an earlier count rule (`≤5,
   wait for impressions`) is not re-planned and never "the next batch": each cell is re-gated
   on its own SERP verdict, and every cell that passes is a `to build` row for step ⓪ of the
-  next session (§10). tabledi's B1–B4 — 27 locale tool cells, 15 SERPs bought — sat under
-  "next batch" through three sessions.
+  next session (§10). tabledi's B1–B4 sat under "next batch" through three sessions — and when a
+  session finally counted, only 15 of its cells had a SERP and 6 of those were unbuilt: a plan
+  is not a verified list until each cell has its own verdict; buy the missing ones ($0.02 each)
+  in the same session, never "next time".
 
 **Page batches — the second table in the ledger.** Keyword rows hold bets; the batch table holds
 what the site's composition is doing, and it is what the batch gate and the retire rule (§7.3)
@@ -295,8 +297,8 @@ run* — all six have been observed in real runs that read as finished.
    An opportunity written in prose (a cluster, a family, a locale axis) with no rows behind it
    is test 1 wearing a different coat. The deliverable counts shipped
    URLs, not qualified rows; a ledger of twenty rows and two pages is a session that spent itself
-   on the analysis. Observed on our own tenant zero: eleven weeks of rounds that produced ledgers
-   and no pages. If a page costs a session, the first deliverable is the generator (§7).
+   on the analysis. Observed on kunavo: five rounds, five pages, three of the rounds at zero; and on clize.ai
+   the opposite failure — eleven weeks in which no round ran at all. If a page costs a session, the first deliverable is the generator (§7).
 
 ---
 
@@ -415,10 +417,11 @@ So: a `no_volume` row whose SERP is already full of category roundups is worth a
 A `no_data` row goes to `seo serp` like any other survivor — the SERP does not need a volume
 figure to show who occupies it. Open and servable → buildable, enrolled as `pending`; wall →
 skip row with the reason. What `no_data` never is: a reason to drop the cell. Measured across
-three sites, the rows with the best positions were disproportionately ones the volume table
-could not see (kunavo's payment long-tails at 4.5 and 5.8, clize.ai's German MCP cluster at
-130 impressions), while the priced `attackable` rows sat at 65+ — the table's coverage and the
-incumbents' coverage are the same coverage. On 2026-09-02, 82 of 119 cells left the funnel at
+three sites: the demand the table could not see was real (clize.ai's German MCP cluster,
+`no_data` in the table, 130 impressions in Search Console), and the only page-1 positions any
+of the sites hold are on such cells (kunavo's payment long-tails at 4.5 and 5.8), while the
+priced `attackable` rows sat at 65+ or unseen — the table's coverage and the incumbents'
+coverage are the same coverage. On 2026-09-02, 82 of 119 cells left the funnel at
 the pricing step without a single SERP look; keeping them would have cost $1.64 in serp calls.
 
 ### `--ai` is a different unit
@@ -547,16 +550,27 @@ a family generator existed, and tabledi shipped forty in two days from one. Buil
 before the second page of any family, and never machine-translate a family page.
 
 **On a site that already has tools, this round's page count is the locale siblings of the tools
-it has.** A sibling is one strings entry (tabledi: a `DICT` table in `tools/apps.js`; clize.ai:
-a `strings.<loc>.mjs`) and its gate is its own SERP (§7.2). Count existing tools × verified
-locale cells before anything else — tabledi had 23 tools, 6 markets, 27 cells already verified
-open, and three sessions wrote "next batch" beside them while building one page. New tools are
-code and tests, one to three per session: they are the next batch's rows, not this batch's
-excuse. And **building a page that passed the gates is never a question for the human.** The
-human's gates are money, external-facing changes (a redirect, DNS, a published post) and sends.
-Ask for those in one line and build while you wait — never bundle the pages into the question
-(tabledi, 2026-09-02: one redirect that needed a yes was asked together with three pages that
-did not, and all four waited).
+it has**, and the gate is each cell's own SERP (§7.2). Siblings come in two prices. A *format*
+tool (viewer, merge, split, convert) is one strings entry (tabledi: a `DICT` table in
+`tools/apps.js`; clize.ai: a `strings.<loc>.mjs`) — ten in a session is ordinary. A *text* tool
+(duplicates, compare, clean) carries locale semantics — full-width and half-width forms in
+Japanese, composed and decomposed accents in French — and a sibling that ignores them reports
+"0 duplicates" on a real file, which is gate 3 failed; budget two or three of those per session.
+Count existing tools × verified locale cells before anything else, and buy the SERPs the cells
+lack ($0.02 each) in the same breath — tabledi had 23 tools and 6 markets; when a session
+finally counted, 15 cells had a SERP, 6 were passed and unbuilt, and "next batch" had been
+written beside them three times. New tools are code and tests, one to three per session: the
+next batch's rows, not this batch's excuse.
+
+**Building and publishing a page that passed the gates on the user's own site is never a
+question for the human.** The human's gates are money, changes that alter what the world already
+sees (a redirect, DNS, a canonical), sends, and posts on other people's platforms. Ask for those
+in one line and build while you wait — never bundle the pages into the question. When a
+human-gated change shares a deploy with pages (a redirect map in the same `vercel.json`), split
+the deploy: ship the pages, hold the change. Step ⓪ is done when the pages are *live*, not
+committed — tabledi, 2026-09-02: one session built three things and ended asking about a
+redirect; the next found all three returning 404 in production and ended the same way. Four
+pages built across two sessions, zero live. A session does not end with built, undeployed pages.
 
 ### 7.1 The pipeline — five steps, each with a pass/fail
 
@@ -586,11 +600,17 @@ Judgment calls that matter:
   translates almost for free, and the non-English cells of your matrix are the low-competition
   ones. Each locale page is a real, distinct intent — this is not page-stuffing. Build every
   locale cell whose *own* SERP is not a wall (`seo serp --locale`, $0.02); the English page's
-  impressions are not the gate, because locale SERPs are only loosely coupled — a browser CSV
-  tool holds #1 in German where the English SERP is a tutorial wall, and a Korean payment
-  long-tail ranked where its English sibling never did. What does not multiply: marketing and
-  landing pages. Locale copies of those earned zero on every site measured (72 on clize.ai, 199
-  near-identical on kunavo) and are the composition drag gate 4 exists to prevent.
+  impressions are not the gate, because the same intent draws a different page in each market —
+  kunavo's `claude code pricing`: `official_wall` in de-DE and es-ES, `open` in ja-JP and
+  fr-FR; one keyword, four verdicts. What does not multiply: marketing and landing pages. Read
+  that evidence for what it varied, not for what it sounds like: clize.ai's 72 translated
+  landing pages earn about what their English originals earn — impressions in the tens, zero
+  clicks — so the failure is the *type*, in every language; kunavo's 199 untranslated copies
+  were duplicates, collapsed by canonicalization, and say nothing about locales; the locale
+  pages that do earn on both sites are translated guides and tools (kunavo's Korean pricing
+  guide is its top non-brand page). The bar for adding a locale to a marketing page is local
+  non-brand impressions on the ones that exist — indexing is the floor, not the pass. Locale is
+  not a page-generation primitive; a SERP-backed intent is.
 - **Pick tools with a shelf life.** Protocol- and format-level tools (llms.txt, schema, robots)
   keep earning links for years; trend-chasing tools die with the trend. Ask: will anyone search
   this in a year?
@@ -622,16 +642,23 @@ do the protecting, not experience:
      locale page written in-locale, a host page whose install and config really differ;
    - **the last batch is being indexed** — `gsc.pages.unseen` (§9) shows the previous batch's
      `index`; find the batch by `firstSeen`, and book the counts in the ledger's batch table (§3).
-     Mostly `indexed` → ship the next family. Mostly `Crawled - currently not indexed`
-     → that is Google's own verdict on those pages; the round is indexing and pruning, not
-     shipping.
+     Read the four states apart: `URL is unknown to Google` and `Discovered - currently not
+     indexed` are the queue (a day-old batch sits there; not a verdict); `Crawled - currently
+     not indexed` is Google's verdict on a page it has read — after you rule out your own hand
+     (a canonical pointing elsewhere, hreflang missing or one-way, a stray `noindex`), which
+     produces the same state. Mostly indexed → ship the next family. Mostly crawled-not-indexed
+     on a clean implementation → the round is indexing and pruning, not shipping. This check
+     paces shipping; it does not certify composition — kunavo's 440 locale model pages indexed
+     fine and earned nothing. Composition is the second check and the retire rule.
    Impressions are not the gate for the next batch: at zero authority they arrive five to six
    weeks after shipping and clicks may not arrive in a quarter, so a rule that waits for them
    ships nothing for months — that was the mechanism behind five pages a month. Waiting is for
    *deciding a new page type* (§10), never for the siblings of a type already decided. The
    retire rule is the other half of shipping fast: a page indexed for two windows with no
-   impressions, or refused by the index for two, is noindexed, merged or redirected. The
-   composition stays clean only if dead pages leave.
+   impressions, or refused by the index for two, is retired: `noindex` by default (the page keeps
+   serving users, only its index exposure goes), merge or redirect when another page carries the
+   intent, delete only what nobody visits. The composition stays clean only if dead pages leave
+   the index.
 
 ### 7.4 E-E-A-T floor and locales
 
@@ -891,8 +918,9 @@ every row the round touched, one line each:
 it), `skip` (with the reason). A `to build` row carries a blocker, and only three exist:
 `needs human` (a redirect, a send, money — asked in one line, §7), `needs tool code` (a new
 client-side tool; name the session it ships in), `needs generator` (no family scaffold yet).
-"Analysis round", "next batch" and "awaiting go-ahead" are not blockers — a row wearing one of
-those means the session is not finished, and it goes back to step ⓪. Second, **placement** — target /
+"Analysis round", "next batch", "awaiting go-ahead" and "needs a SERP" are not blockers — a
+SERP is $0.02 inside the fuse and is bought on the spot — and a row wearing one of those means
+the session is not finished: it goes back to step ⓪. Second, **placement** — target /
 page / status (`sent` / `drafted` / `blocked: needs human`). Then at most five lines of
 narrative. A report that is findings and no pages table is test 1 (§3) in its final form: the
 human cannot act on a paragraph. A round with an empty pages table says so *in* the table, with
@@ -925,6 +953,14 @@ So: **promise time series against yourself, never absolute share.** "Our AI refe
 it. Breaking this rule in customer-facing copy is a product-integrity failure, not a rounding
 error.
 
+**The cases in this file are samples of one to three sites, and each varied one thing.** Before
+a case is cited for a rule, say what it varied, how many, and over which window: clize.ai's 72
+locale pages varied *page type* (their English originals earn the same nothing); kunavo's 199
+copies varied *translation* (identical, so Google collapsed them); tabledi's page-1 tools varied
+*age* (three years), not links; tabledi's 36 marketing locales were a day old when first read.
+A rule that survives only because its example was misread is a rule waiting to fail — three
+such misreadings were found in this file on 2026-09-02 and corrected.
+
 ---
 
 ## What this skill will not do