Fix what four cold-eyes reviewers found, and write the lesson into the skills

[pftg]

Fix what four cold-eyes reviewers found, and write the lesson into the skills

Paul asked for a cold-eyes re-review instead of a script. Four reviewers - ICP
cold read, voice and AI tells, competitor standard, claim verification - found
real defects in all three posts. Every category below is something my own checks
passed.

Numbers, in the post whose thesis is "say the number"

• The "before" lychee fence had 15642 Unique. The recorded run says 31937.
  A tool output reproduced inside a fence with one field silently edited.
• [snap_diff] 287 screenshots compared - 287 is the unit-test run count
  (.okf/log.md:229). Real snap_diff records say 55. Two different numbers
  merged from memory.
• 114,239 / 31,888 / 149,740 / 35,501 came from MY runs today; the fix landed
  at 114,050, which is what .okf/log.md, STATUS.md and the design-system
  README all record. A fresh number used to illustrate a historical event.
• "reported zero broken links for months" - the gate shipped 2026-07-21, one
  month before. And the repo's own Rakefile:100 and test-gates.md:850 said
  "for a year", which git history contradicts. I softened a wrong number instead
  of checking it. Both source records corrected here too.

Sourcing

• "Both publish their error bars" - Pew publishes none. Graphite publishes
  4.2% and flags its own finding as superseded. Rewritten to say what each
  actually offers, because the difference is the interesting part.
• "eleven of twelve" - GPTZero flagged all twelve (11 high confidence, one
  moderate). The phrasing implied one was clean.
• "$900,000" - Piro received $900,000 for its work; the post said it was spent
  on the fake think tank.
• "Responsible Statecraft traced it" - NewsGuard analyst Alice Lee traced it;
  RS reported it. Verifying a quote is not verifying the sentence around it.
• The opening ```diff was paraphrased while presented as literal. Replaced
  with both real lines from 7de032e24.
• The blockquoted reviewer verdict appears in no log. Reconstructed from memory
  and presented as a quotation - now stated as reported substance, unquoted.

Commands that did not run

for f in content/blog/**/index.md matches nothing in bash without
shopt -s globstar - it reports success having checked zero files, in a post
about checks that check nothing. Switched to find, and the post now says so.

The anonymous-subject grep was shipped without a hit count in a post that
explains why unmeasured patterns cry wolf - and it matches "a large number of",
"a large team". Removed, with the reason.

Structure

• All three opened on the identical two-beat reveal. Posts 1 and 2 reshaped;
  post 3 keeps it, where it lands hardest.
• The Laravel anecdote carried two same-day posts. Post 1 keeps the detailed
  version; post 2 states the mechanic without a second proof-signal.
• Post 1's prose said "only one of them is automatable" while its own diagram
  caption said "only the first two" - they contradicted each other. Fixed.
• SQLite was on post 3's cover chip and in its Sources, and the body never told
  the story. Now told, where it belongs.
• HN comment counts dropped from Sources; §6b of the voice guide names them as a
  non-receipt.
• Rašić was cited and never appeared in the body. Now engaged.

The ratchet had the same bug I keep finding

It fired on the deleted line inside my own ```diff - the phrase appears there as
a quoted exhibit of bad writing, not as our claim. Fenced code is now stripped
before the phrase scan, same reasoning as the heading markers. Fix the
instrument when the instrument is wrong.

Skills

blog-write gains what the panel found and how to run it: four lenses, spawned
as a DIFFERENT agent type than the writer, briefed with goal and artifact,
each required to name something it would cut, run in parallel and all four read
before editing. blog-operator gains the sequencing consequence - a post is
draft plus panel plus fix pass, not one unit.

Gates: marketing_copy_test 5 runs / 13 assertions / 0 failures. bin/hugo-build
green. bin/rake test:links clean at 31,948 unique links, zero errors.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_016PUkwFTsiv7EB2DYKogbpg

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: fdb8be18-2f85-4801-bbdc-ce75631e37d2

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}