Add /blog-next: the pipeline had no owner for what to write - #580

Topic selection had no owner. Any request to write a post entered the pipeline at the draft step, so a decayed queue row or a dedup collision was only discovered after a draft existed. This adds the missing half and records what three live GSC pulls did to the queue.

Docs, skills and .okf/ only — no content/, no templates, no CSS. Per #577, docs-only merges on local gates.

Per commit

Add the /blog-next skill — Stage A (generate candidates, then audit them), Stage B (research), Stage C (outline + gate). Never writes a post; hands a gated outline to blog-post-coordinator.

The load-bearing rule, learned from two audits that falsified the queue's own lead item: search demand cannot generate topics on this property, it can only veto them. The blog earns ~0.69 named non-brand clicks/day and the ICP is absent from the query data, so a search-driven queue returns HOLD forever.

.gitignore needed three fixes for the skill to ship at all: the directory was ignored as claude-flow scaffolding; a directory negation does not cover file-pattern rules matching deeper; and .claude/**/*.json still swallowed evals.json inside the un-ignored directory. Caught by peer review, not by me — git check-ignore -v returns exit 0 on an un-ignored file, so it reads as confirmation either way. git status --untracked-files=all is the reliable check.

Wire /blog-next into the pipeline — both blog agents now accept a gated outline instead of re-deriving the topic; the batch orchestrator reads §13 (§12 was retired but still referenced).

Adds STEP 4e — SELF-CRITIQUE (BLOCKING): reflexion-reflect before the critic panel. Already live in the LinkedIn and course lanes, simply absent from the blog lane — the lane with the most voice-sensitive output.

Fixes a gate-inverting collision in CLAUDE.md: slop >= 8/10 (blog, higher better) and Slop <= 25 (course, 0-100, lower better) were both written as "the slop gate".

Content can run unattended, but only with three exits — SHIPPED / HOLD-with-evidence (a terminal success) / BLOCKED on a named decision. A two-exit loop always ships, because stopping feels like failing.

Not theoretical: two /blog-next runs this day correctly ended in HOLD. A two-exit loop writes both posts. The completion promise must be about gates, never quality — a promise containing "the post is good" is unfalsifiable, and an agent under loop pressure will assert it.

Three live GSC pulls — §13d/e/f, each recording the number that killed a row:

• N2 falsified. Its 52x is a denominator, not a reader response: (c1/i1)/(c2/i2) factorises to 1.4x clicks × 37.2x impressions. N3's addressable volume corrected 6,895 → ~550 (synthetic query family).
• §13d's own replacement retracted — one query, already won, ceiling +0.13 clicks/day.
• N7 "how AI helps developers" DO-NOT-WRITE — 61 slugs already match, six of them shipped near-duplicate pairs; 180 query rows at zero clicks; and it addresses developers where the ICP is a non-technical founder.
• N8 vertical ai agents FLAG-not-write — the first row whose arithmetic clears the demand floor, and still not worth a slot: family decaying 2x, SERP held by Salesforce/IBM/Google, our asset a 3-minute dev.to import.
• N9 SCHEDULED — first-party benchmark of the AI code-search MCP tools.

analytics-access.md gains the traps these pulls earned: Trap C (page position is an artifact whenever the CTR is), Trap D (query-vs-query is the legitimate comparison level), the synthetic-query-family tell, and Trap A generalised — a filter truncates a denominator as readily as a row limit, and has_more: false means complete given the filter. Both generalisations came from my own errors this session, documented in §13f.

Rewrite bip-agents post idea-first — opened with a parable and reached the idea in beat 3; influencers in this lane open with the claim. 141 words, zero em dashes. reflexion-reflect caught three defects self-review had passed: a slogany flip, "explicitly rejected" where the notes say "deferred", and an unverifiable "The same morning".

§13g — Paul's "structure a team with AI harnesses" row resolves to UPGRADE claude-code-xp-team-workflow, not a new post. Stated in the section: that verdict rests on reading the post, not a measurement chain, so it did not pass the Stage A reviewer gate the other verdicts did — flagged so nobody later cites it as gate-passed.

Gates

• bin/hugo-build — green, 1181 pages, no errors.
• Visual suites correctly skipped: the diff touches no template, stylesheet or body HTML.

Review links

• LinkedIn draft: http://localhost:<session port>/linkedin/personal/bip-agents-cancelled-own-backlog/

🤖 Generated with Claude Code

https://claude.ai/code/session_016PUkwFTsiv7EB2DYKogbpg