The Astra/Fable
SEO Folder

---
name: best-page-analyser
description: Rank the site's pages by opportunity, not by clicks, and give each one a verdict. Run first, and again whenever gsc-data.md is refreshed.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing if it is not in gsc-data.md:
- the PAGES section, with both periods
- the BRAND TERMS list

PROCESS
1. Drop every page under the page floor in scoring.md.
2. Compute the non-brand average CTR per position band from the QUERIES section.
   Write the three numbers out. Tell me to paste them into BAND AVERAGES in
   gsc-data.md so the other skills can use them.
3. For every remaining page, compute opportunity using scoring.md. Show the
   three inputs (impressions, ctr, target_ctr) next to it.
4. Give each page one verdict:
   DEFEND   winning band, CTR at or above its band average
   PUSH     striking distance or page two, opportunity above 20 clicks
   RESCUE   clicks down 30%+ versus the previous period (hand to content-decay-tracker)
   IGNORE   noise band, or opportunity under 20 clicks
5. Sort by opportunity, descending.

OUTPUT — one table, capped at 10 rows:
url | position | impressions | ctr | band avg ctr | opportunity | verdict

RULES
- Opportunity is only computed with the formula in scoring.md. Never estimate
  a "potential traffic" figure any other way.
- A page whose position is in the winning band gets opportunity 0, even if its
  CTR looks low. Say "defend" and move on.
- If BAND AVERAGES cannot be computed because QUERIES is empty, say so and stop.

End with ONE line: the single PUSH page to open first, and why in ten words.

---
name: keyword-opportunity-finder
description: Find the queries the site already gets impressions for but does not win, and say which page should own each one. Use before any content planning.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the QUERIES section, both periods
- the PAGES section (to map queries to existing URLs)
- BRAND TERMS

PROCESS
1. Drop brand queries and every query under the query floor.
2. Keep queries with average position between 8.0 and 30.0. Above 8 is
   already winning enough; below 30 is not yet a signal.
3. Compute opportunity for each with scoring.md.
4. For each query, find the existing URL most likely ranking for it: the page
   whose topic matches, using QUERY x PAGE if present, otherwise your best
   reading of the PAGES list. Mark that reading "probable" if you had to guess.
5. Decide: STRENGTHEN the existing page, or NEW PAGE if no existing page is
   a credible home for the query.
6. Group queries that a single page could satisfy. One group, one destination.

OUTPUT — one table, capped at 15 rows, sorted by opportunity:
query | position | impressions | opportunity | destination URL or "new" | action

RULES
- Never invent search volume. Impressions are the only demand number allowed.
- No difficulty score. You cannot see one, and inventing one is worse than
  leaving it out.
- A query maps to exactly one destination. If it fits two pages, that is a
  cannibalization-detector job. Say so and pick the higher-impression page.

End with ONE line: the single query to go after this month, and its destination.

---
name: search-intent-classifier
description: Read a live SERP and name the intent class and the page format that can rank. Use before writing or rewriting any page.
---

Read scoring.md first.

INPUTS — ask for what is missing:
- the query, and the country and language it should be read in
- the SERP itself: a screenshot, a paste of the top 10, or a live fetch
If you cannot see the SERP, say so and stop. Do not describe what the SERP
"probably" looks like. That is the one failure mode of this skill.

INTENT CLASSES
  informational  guides, forums, videos, People Also Ask dominate
  comparison     listicles, "best", "vs", review sites dominate
  transactional  product pages, pricing, ads above the fold dominate
  navigational   one brand owns the top 3. Never target it.

PROCESS
1. List the top 10 results: URL, page type, and whether it is a brand's own page.
2. Count the page types. The majority type is the format Google is rewarding.
3. Note SERP features present: ads, People Also Ask, video, images, local
   pack, shopping, AI overview.
4. Assign the intent class.
5. Judge whether a page of the winning format is realistic for this site.

OUTPUT
- intent class, one word
- winning format, one line, with the count that justifies it ("7 of 10 are listicles")
- SERP features present, comma separated
- what would have to be true for us to rank, 2 bullets maximum
- verdict: TARGET or SKIP, with the reason in one sentence

RULES
- Only describe results you can actually see. Never reconstruct a SERP from
  memory. Rankings change and yours are months stale.
- If the top 3 are one brand's own properties, the verdict is SKIP.
- Hand the intent class to content-brief-builder. Never write the page from here.

End with ONE line: TARGET or SKIP, and the format if TARGET.

---
name: quick-win-detector
description: Find the pages that meet the quick-win definition in scoring.md and name the specific fix for each. Use weekly.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the PAGES section with BAND AVERAGES filled in
- QUERY x PAGE for the top pages. If it is empty, say "query-to-page pairs
  missing: fix type will be limited to title/meta" and continue on PAGES only.

PROCESS
1. Keep pages with average position 4.0 to 15.0 and above the page floor.
2. Compute opportunity with scoring.md. Drop anything under 20 clicks.
3. For each remaining page, diagnose the cheapest fix from the data:
   TITLE/META   CTR below band average, position stable -> effort 1
   INTRO        top query's words absent from the page's title and first 100
                words (only if the page text was provided) -> effort 2
   LINKS        position 8-15 with impressions rising, few internal links
                (only if a crawl or link count was provided) -> effort 2
   SCHEMA       page type matches a rich-result type and has no schema
                (only if the page source was provided) -> effort 2
   If you cannot tell which, write "fix: unknown, open the page" and effort 2.
4. Drop anything that needed effort 3 or 4. That is not a quick win.
5. Rank by priority = opportunity / effort.

OUTPUT — one table, capped at 7 rows:
url | position | impressions | ctr | band avg ctr | opportunity | fix | effort | priority

RULES
- Never label a winning-band page a quick win. It is a page to defend.
- Every "fix" must be justified by a number in gsc-data.md or a file the
  user provided. No "the content could be better".
- If nothing meets the definition, say "no quick wins this week" and stop.
  A short list is a valid output. An inflated one is not.

End with ONE line: the single quick win to ship today, and the fix.

---
name: topical-authority-mapper
description: Build the site's topic map from its own query data, mark coverage per cluster, and name the missing page that matters most. Use monthly, before briefing new content.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the QUERIES section (non-brand)
- the PAGES section
- optionally, the output of keyword-opportunity-finder

PROCESS
1. Drop brand queries and queries under the query floor.
2. Group the remaining queries into topics: a PILLAR is a broad head term the
   site gets impressions for; a CLUSTER is a group of 3+ queries under it that
   one page could satisfy. Fewer than 3 queries: fold into the nearest cluster.
3. Cap at 5 pillars, chosen by total impressions.
4. For each cluster, find the URL that currently ranks for most of its queries
   and mark coverage:
   COVERED  a page exists and ranks in the top 10 for the cluster's top query
   THIN     a page exists but ranks 11+ for the top query
   MISSING  no page in PAGES plausibly targets the cluster
5. Sum the impressions of every MISSING and THIN cluster per pillar.

OUTPUT
For each pillar (max 5):
  pillar | total impressions | clusters covered / thin / missing
Then one table of the MISSING and THIN clusters, capped at 10 rows, sorted by impressions:
  cluster (named after its top query) | pillar | impressions | coverage | current URL or "none"

RULES
- The map is built from gsc-data.md only. Do not add topics the site has no
  impressions for; that is competitor-gap-analyser's job.
- One cluster maps to exactly one URL. That constraint is what prevents
  cannibalization later.
- Never state a "topical authority score". There is no such number in the data.

End with ONE line: the single MISSING cluster to brief first, and its pillar.

---
name: content-brief-builder
description: Turn one cluster and one intent verdict into a one-page brief a writer or draft-writer can execute. Use once per new or rewritten page.
---

Read gsc-data.md and scoring.md first.

INPUTS — ask for what is missing:
- the target cluster: its top query and the 3-10 queries under it
- the intent class and winning format, from search-intent-classifier
- the destination URL, existing or "new", and the pillar it belongs to
- optionally: the top 3 ranking pages' headings, if the user pasted them

PROCESS
1. State the one job of the page in a sentence: who searches this and what
   they need to leave with.
2. Write the H1 and an outline of H2s. The outline follows the winning format,
   not a generic template. A comparison page gets a comparison outline.
3. List the questions the page must answer, drawn from the cluster's queries.
   One question per query, deduplicated.
4. List the entities and terms the page must mention, only ones present in the
   cluster's queries or the pasted competitor headings.
5. List internal links: 3-5 pages from PAGES that should link TO this page,
   with the anchor text (a query the page targets), and 2-3 this page should
   link OUT to.
6. Give 3 title candidates under 60 characters and one meta description under
   155 characters.
7. If competitor headings were provided, give a length range from them.
   If not, write "length: not observed".

OUTPUT — under 400 words, in this order:
job | H1 | outline | questions | terms | internal links in | internal links out |
titles | meta | length

RULES
- No word counts, keyword densities or "LSI keywords" pulled from nowhere.
- Every internal link source must be a URL in gsc-data.md.
- Do not write body copy. That is draft-writer's job.

End with ONE line: the single section that must be the best on the web for
this page to rank.

---
name: title-meta-rewriter
description: Rewrite one page's title and meta description from its own query data, and pick one. Use on any page flagged TITLE/META by quick-win-detector.
---

Read gsc-data.md and scoring.md first.

INPUTS — ask for what is missing:
- the URL
- its current title and meta description (paste them; do not fetch from memory)
- its queries from QUERY x PAGE, or the top QUERIES you can attribute to it
- what the page actually contains, in one sentence from the user

PROCESS
1. Name the top query by impressions and the CTR gap versus its band average.
2. Diagnose in one sentence why the current title underperforms: query words
   missing, promise unclear, truncated, or duplicate of another page's title.
3. Write 3 titles, each under 60 characters, each starting with or containing
   the top query's words in natural order. One plain, one with a specific
   benefit the page delivers, one with a number that is ACTUALLY on the page.
4. Write 1 meta description under 155 characters that says what the reader
   gets and ends with a verb.
5. Pick ONE title. Say why in one sentence, using the CTR gap.

OUTPUT
diagnosis | title A | title B | title C | meta | chosen: A/B/C, reason

RULES
- No number, year or claim in a title that the user has not confirmed is on
  the page. "2026 guide" on a page that says nothing about 2026 is banned.
- No clickbait patterns: "you won't believe", "the secret", "ultimate".
- Do not change the title of a winning-band page unless its CTR is below its
  band average. Say "defend, do not touch" instead.
- Character counts are shown next to each candidate.

End with ONE line: the chosen title, ready to paste.

---
name: internal-link-builder
description: Propose internal links to one target page, with source, anchor and placement, from pages that already have traffic. Use on any page flagged LINKS, or on any new page after it is published.
---

Read gsc-data.md and scoring.md first.

INPUTS — ask for what is missing:
- the target URL and its top 3 queries
- the PAGES section (candidate sources)
- optionally: the text of candidate source pages, or a crawl export with
  existing internal links

PROCESS
1. Candidate sources are pages in PAGES above the page floor. Rank them by
   topical closeness to the target's top query: shared words, same pillar.
2. Drop any source that already links to the target, if a crawl was provided.
3. For each source, choose an anchor: one of the target's top 3 queries, or a
   close natural variant of it. Never the bare URL, never "click here".
4. For each, write the sentence the link belongs in. If the source text was
   provided, quote the existing sentence to edit. If not, write a sentence to
   add and mark it "new sentence".
5. Cap at 8 links. Fewer is fine.

OUTPUT — one table:
source URL | source impressions | anchor text | placement sentence | existing or new sentence

RULES
- No site-wide links: no nav, footer or sidebar suggestions. Body links only.
- Never suggest a source page that is not in gsc-data.md.
- Two links from the same source to the same target is one link. Do not pad.
- If the target's queries are not in the file, say so and stop. You cannot
  choose an anchor without them.

End with ONE line: the single link to add first, source and anchor.

---
name: cannibalization-detector
description: Find queries where two or more of the site's pages compete and neither wins, and say what to do with each pair. Use monthly.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the QUERY x PAGE section. Without it this skill cannot run. Say
  "query-to-page pairs missing, export queries per page for the top 10 pages"
  and stop. Do not infer pairs from the PAGES list.

PROCESS
1. For every query that appears under 2 or more URLs, list the URLs with
   position and impressions.
2. Apply the definition in scoring.md: keep only pairs where both URLs are in
   the top 20 and neither is above position 5.
3. Drop queries under the query floor.
4. For each remaining conflict, choose a verdict:
   MERGE         same intent, same audience: fold the weaker into the stronger, redirect
   DIFFERENTIATE different intent (one informational, one transactional): retitle
                 the weaker page toward its own intent
   REDIRECT      the weaker page has no queries of its own
   LEAVE         positions are moving apart month over month; check again next month
5. Rank by the query's impressions.

OUTPUT — one table, capped at 10 rows:
query | impressions | URL A (position) | URL B (position) | verdict | which page survives

RULES
- The surviving page is the one with more clicks across ALL its queries, not
  the one with the better position on this one query. Show both numbers.
- Never recommend deleting a page without a redirect target.
- If no pair meets the definition, say "no cannibalization by this definition"
  and stop. Do not lower the bar to produce a list.

End with ONE line: the single conflict to resolve first, and its verdict.

---
name: content-decay-tracker
description: Find the pages losing clicks while demand holds, separate decay from demand change, and say which to refresh first. Use monthly.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the PAGES section with BOTH periods. One period is not enough to detect
  anything. Say so and stop.
- the CHANGELOG (to rule out changes we made ourselves)

PROCESS
1. For every page above the page floor, compute clicks change and impressions
   change versus the previous period, in percent, with the raw numbers.
2. Apply the definition in scoring.md:
   DECAY          clicks down 30%+, impressions flat or down less than 10%
   DEMAND CHANGE  clicks down 30%+ AND impressions down 10%+
   Everything else is not this skill's business.
3. Check the CHANGELOG. If we changed the page inside the window, label the
   decay "self-inflicted, probable" and name the change.
4. Check seasonality per scoring.md. If last year's data is not in the file,
   label every finding "seasonality unchecked".
5. For each DECAY page, recommend one of:
   REFRESH      position slipped 2+ places: update the content, re-check intent
   RETITLE      position held, CTR fell: hand to title-meta-rewriter
   CONSOLIDATE  a newer page of ours now ranks for its queries: hand to
                cannibalization-detector
6. Rank by clicks lost.

OUTPUT — one table, capped at 10 rows:
url | clicks now | clicks before | change | impressions change | position change | label | recommendation

RULES
- Every percentage is shown with the two raw numbers behind it.
- Demand-change pages are listed separately, in 3 rows maximum, with no
  recommendation. There is nothing to rewrite.
- Never attribute a drop to an algorithm update. You cannot see one in this data.

End with ONE line: the single page to refresh first, and the recommendation.

---
name: competitor-gap-analyser
description: Compare a competitor's ranking topics with the site's own coverage and name the gaps worth a page. Use quarterly, or when planning new content.
---

Read gsc-data.md and scoring.md first.

INPUTS — ask for what is missing:
- one competitor domain
- that competitor's top ranking pages or queries, pasted from a tool export or
  from SERPs the user looked at. If none is provided, say "I cannot see a
  competitor's rankings; paste their top pages or queries" and stop. Do not
  describe what a competitor "probably" ranks for.
- the output of topical-authority-mapper, if it has run

PROCESS
1. Group the competitor's pages or queries into topics, the same way
   topical-authority-mapper does.
2. For each topic, check our coverage in PAGES and QUERIES:
   NO PAGE     nothing in PAGES plausibly targets it, and we have no impressions
   WEAK        we have impressions but average position 20+
   COVERED     we rank in the top 20. Not a gap. Drop it.
3. For each NO PAGE or WEAK topic, judge fit: is this topic something our
   site should credibly own, given the pillars we already have? YES / NO.
4. Where the competitor's demand number is present in the paste, show it and
   label it "competitor tool estimate". Where it is not, write "demand: not in data".
5. Rank by fit first, then by our impressions if WEAK, then by the estimate.

OUTPUT — one table, capped at 10 rows:
topic | competitor URL | our coverage | our impressions or "none" | fit | verdict: BRIEF / SKIP

RULES
- Never merge a competitor's estimate with our GSC numbers in one column.
  They are not the same unit.
- A gap with fit NO is listed once with SKIP and no further commentary.
- Do not recommend copying the competitor's page. Hand BRIEF topics to
  search-intent-classifier, then content-brief-builder.

End with ONE line: the single gap to brief first, and why it fits.

---
name: schema-generator
description: Generate one valid JSON-LD block for one page, of the type that matches the page, using only facts present on it. Use on pages flagged SCHEMA, or on any new page before publishing.
---

Read scoring.md first (for the one-action rule).

INPUTS — ask for what is missing:
- the URL
- the page's content, pasted or fetched. If you cannot see the page, say so
  and stop. Schema written from a URL alone is invented.
- whether the page already has schema (paste it if so)

PROCESS
1. Name the page type from its content, and pick ONE schema type:
   Article / BlogPosting   an editorial page with an author and a date
   Product                 a product with a name, and a price if shown
   FAQPage                 a page with visible question-and-answer pairs
   HowTo                   a page with numbered steps the reader performs
   Organization            the home or about page only
   BreadcrumbList          add alongside any type above, if the breadcrumb is visible
   If the page matches none, say "no schema type fits this page" and stop.
2. Fill the block only with values that are visibly on the page: names, dates,
   prices, questions, steps. Copy them exactly.
3. If the page already has schema, say whether to replace or merge, and why.
4. Validate: every required property present, valid JSON, no empty strings.

OUTPUT
- schema type, and the one sentence that justifies it
- the JSON-LD block, ready to paste inside a script tag of type application/ld+json
- where to paste it: in the page head, or in the template if the CMS handles it
- the one property you left out and why, if any was required but not on the page

RULES
- No aggregateRating, review, or ratingValue unless the page visibly shows
  reviews with a count. Invented ratings are a manual-action risk.
- No FAQPage for questions that do not appear on the page as questions.
- No dates, authors or prices that are not on the page.
- One type. Do not stack Article and FAQPage and HowTo on one page.

End with ONE line: the type, and "paste into head" or "paste into template".

---
name: page-priority-ranker
description: Merge everything the audit skills found into one priority list scored by opportunity over effort, and name the page for this week. Run after best-page-analyser, quick-win-detector, cannibalization-detector and content-decay-tracker.
---

Read gsc-data.md and scoring.md first.

INPUTS — use what is present, and say which are missing:
- the outputs of best-page-analyser, quick-win-detector, cannibalization-
  detector and content-decay-tracker, pasted or from earlier in this session
- if none are present, run on PAGES alone with the opportunity formula and
  effort 2 for every row, and label every effort "assumed"

PROCESS
1. Collect every page named by any input skill into one list. A page named
   twice is one row; keep the highest opportunity and note both sources.
2. For each row, carry over or compute: opportunity (scoring.md formula),
   effort (from the fix type the source skill named), priority = opportunity
   / effort, confidence label.
3. Attach the NEXT SKILL for the row:
   TITLE/META fix    -> title-meta-rewriter
   LINKS fix         -> internal-link-builder
   SCHEMA fix        -> schema-generator
   DECAY refresh     -> content-brief-builder, then draft-writer
   CANNIBALIZATION   -> the verdict from cannibalization-detector
   DEFEND            -> none, remove from the list
4. Sort by priority, descending. Ties go to the lower effort, then to the
   higher confidence.
5. Cap at 10 rows.

OUTPUT — one table:
rank | url | source skill | opportunity | effort | priority | confidence | next skill

RULES
- Every opportunity number was computed by scoring.md or carried from a skill
  that did. Never adjust a number "to reflect strategic importance".
- Confidence is shown per row. A "probable" row with a huge opportunity does
  not outrank a "confirmed" row unless its priority is at least double.
- Do not add pages from general knowledge of "what usually matters".

End with ONE line: the single page to work on this week, the next skill to
run on it, and the opportunity in clicks.

---
name: draft-writer
description: Write a first draft from a content-brief-builder brief, in the winning format, with every unverified claim flagged. Use once per brief. Never without a brief.
---

Read scoring.md first (for the universal rules).

INPUTS — stop if missing:
- a brief from content-brief-builder: job, H1, outline, questions, terms,
  internal links, length. If there is no brief, say "run content-brief-builder
  first" and stop. Do not write from a keyword.
- optionally: the user's own notes, examples, numbers or experience to include

PROCESS
1. Follow the outline exactly. Same H2s, same order. If a section cannot be
   written honestly, write the heading and "[SECTION NEEDS INPUT: what]".
2. Answer every question in the brief, each in the section where it belongs.
3. Use the brief's terms naturally. Never repeat a term to hit a count.
4. Place every internal link from the brief in a sentence where it earns its
   place, with the brief's anchor text.
5. Flag, do not fake:
   [SOURCE NEEDED]  any statistic, study or external claim
   [EXPERIENCE]     any place where a real example from the user would replace
                    a generic one
   [VERIFY]         any product fact, price, feature or date
6. Stay within the brief's length range. If the brief said "length: not
   observed", write to the outline and stop when the questions are answered.
7. End the draft with the user's CTA if one was given, otherwise "[CTA]".

OUTPUT
The draft, in markdown, with the H1 and every H2 from the brief.
Then a short list: every flag in the draft, one per line, so the user can
clear them in order.

RULES
- No opening with "In today's fast-paced world" or any sentence about how
  important the topic is. Start with the answer.
- No invented statistics, quotes, case studies or customer names. Flag them.
- Write in the second person unless the brief says otherwise.
- The draft is not finished until every flag is cleared by a human. Say so
  at the top.

End with ONE line: the single flag the user should clear first, because the
page cannot publish without it.

---
name: rank-audit-reporter
description: Write the monthly report from gsc-data.md and the changelog: what moved because of us, what moved without us, what did not work, and the one focus for next month. Use once a month, last.
---

Read gsc-data.md and scoring.md first.

INPUTS — stop and say which is missing:
- the PAGES section with both periods
- the QUERIES section with both periods
- the CHANGELOG. If it is empty, say "no changelog: nothing can be attributed
  this month" and write only sections 2 and 4.

PROCESS
1. Site totals: clicks, impressions, non-brand CTR, versus the previous period,
   with raw numbers.
2. For every CHANGELOG entry inside the window, find its page in PAGES and
   compare both periods. Attribute a rise ONLY if the change date precedes the
   rise and the page's impressions did not also jump. Label each "confirmed",
   "probable" or "timing does not support attribution".
3. Find the 3 biggest movers, up and down, that are NOT in the changelog.
   These moved without us.
4. Find every changelog entry older than 60 days whose page has not moved.
   These did not work. Do not skip this section because it is uncomfortable.
5. Take the top row from page-priority-ranker if it ran this month, otherwise
   the top PUSH page from best-page-analyser, as next month's focus.

OUTPUT — under 300 words, in this order:
TOTALS         2 lines, raw numbers and percentages
MOVED BECAUSE OF US   one line per attributed change, with the confidence label
MOVED WITHOUT US      3 lines maximum
DIDN'T WORK           one line per stale change, or "nothing shipped 60+ days ago"
NEXT           one action, carried from whichever skill named it

RULES
- Every percentage is shown with the raw numbers behind it.
- Never attribute a rise to work whose date does not line up. Write "timing
  does not support attribution".
- No revenue projection from clicks unless the user gave a conversion rate
  and a value per conversion.
- No chart descriptions. Numbers and sentences.

End with ONE line: the single thing next month is about.

---
name: workflow-fix-what-ranks
description: The weekly loop. Find the page closest to more clicks, fix the cheapest thing on it, ship it, log it. Run every Monday.
---

Read gsc-data.md and scoring.md first. If gsc-data.md is older than 35 days,
say so and stop: the loop runs on fresh data or not at all.

STEPS — run each skill, hand its output to the next, do not skip
1. best-page-analyser. Keep its table in the session.
2. quick-win-detector. If it returns "no quick wins this week", go to step 5
   with the top PUSH page from step 1 instead.
3. page-priority-ranker, on the outputs of 1 and 2. Take row 1 only.
4. Run the NEXT SKILL that row names: title-meta-rewriter, internal-link-builder
   or schema-generator. Present its single output.
5. Ask me to ship it. Do not proceed until I say it is live.
6. Write the CHANGELOG line for gsc-data.md:
   today's date | url | what changed | which skill proposed it
   Tell me to paste it.

RULES
- One change per week. If I ask for a second, remind me that the reporter
  cannot attribute two changes on one page in one window, then do it if I insist.
- Never touch a DEFEND page in this workflow.
- Total output across all steps under 600 words. Tables count.

End with ONE line: what shipped, and the date to re-check it (28 days on).

---
name: workflow-build-whats-missing
description: The monthly build. Find the one page the site is missing, confirm the SERP wants it, brief it, draft it, mark it up. Run in the first week of the month.
---

Read gsc-data.md and scoring.md first.

STEPS — run each skill, hand its output to the next
1. keyword-opportunity-finder. Keep the rows marked NEW PAGE.
2. topical-authority-mapper. Keep the MISSING clusters.
3. competitor-gap-analyser, only if I paste a competitor's pages. Otherwise
   say "skipped: no competitor data" and continue.
4. Merge the candidates from 1, 2 and 3 into one list. Rank by impressions,
   with any competitor estimate shown separately. Pick the top 3.
5. search-intent-classifier on each of the 3. I will paste the SERPs, or you
   fetch them if you can. Keep the first one that returns TARGET. If all three
   return SKIP, take the next 3 and repeat once, then stop and say so.
6. content-brief-builder on the TARGET cluster.
7. Stop and show me the brief. Do not draft until I approve it.
8. draft-writer on the approved brief.
9. schema-generator on the draft, if a type fits.
10. Write the CHANGELOG line for gsc-data.md and tell me to paste it when the
    page is live.

RULES
- One page per run. The second-best candidate goes in a "next month" line
  at the end, and nowhere else.
- Never brief a cluster whose SERP you have not seen. Step 5 is not optional.
- If step 1 and step 2 return nothing above the floors, say "the site has no
  data-backed gap this month; grow the existing pages instead" and stop.

End with ONE line: the page being built, its target query, and the flag
count left in the draft.

---
name: workflow-protect-and-report
description: The month-end loop. Catch decay and cannibalization before they cost a quarter, patch the worst one with links, then write the report. Run on the last working day of the month, after gsc-data.md is refreshed.
---

Read gsc-data.md and scoring.md first. Confirm gsc-data.md was exported this
week. If not, stop and ask for a fresh export: a month-end report on
mid-month data is not a report.

STEPS — run each skill, hand its output to the next
1. content-decay-tracker. Keep the DECAY rows.
2. cannibalization-detector. If QUERY x PAGE is empty, it will stop; note
   "cannibalization unchecked" for the report and continue.
3. Take the single worst row across 1 and 2: the DECAY page with the most
   clicks lost, or the cannibalization conflict with the most impressions,
   whichever has the larger number.
4. If it is a DECAY page marked RETITLE or REFRESH: internal-link-builder
   on it. Present the first link only.
   If it is a cannibalization conflict: present the verdict and the surviving
   page. Do not execute a merge; that is a human decision.
5. Ask me to ship the link or approve the verdict. Write the CHANGELOG line.
6. rank-audit-reporter. Its NEXT line becomes next month's focus.

RULES
- Steps 1 and 2 may each return "nothing found". That is a good month, not a
  failure. Say so in the report.
- Never fix more than one thing in this workflow. The other rows are carried
  into the report's DIDN'T WORK or MOVED WITHOUT US sections, not acted on.
- Total output under 700 words including the report.

End with ONE line: next month's single focus, from the report.