kai-gate

Kai root note: knowledge/, harness/, and scripts/ paths in this skill live in the Kai install, not the user's project. Resolve them against the first ancestor directory of this SKILL.md that contains a knowledge/ folder (the Kai plugin root, ~/.claude/kai, or the kai-cmo-harness repo). MARKETING.md, memory/, and any output files live in the current project. If a referenced scripts/ command is not available in this install, say so, skip it, and continue with the file-based guidance — never fabricate its output.

Objective

A verdict on a piece of content that the writer cannot argue with: a per-dimension Four U's score against the right threshold, every banned word and slop phrase located by line, every voice-pattern regex run, and SEO lint applied when the content targets search. The verdict is PASS or FAIL with the specific failing rule named — never a vague quality impression.

Done when

Work type internal-research — floor E2/C2/O0 (harness/eco-floors.yaml). The gate report is an internal artifact; nothing leaves the workspace, and SHIPPED is terminal.

• E2 — the gate report exists and carries every section below: Four U's total plus each dimension with its reason, banned words, AI slop, voice patterns, SEO lint (or SKIPPED), and an overall verdict.
• C2 — every declared check ran at its stated threshold. A skipped check is reported as SKIPPED with the reason, never silently omitted. O0 — no outcome obligation.

This skill grades other people's work. It does not issue a completion verdict on its own gate report, and a gate PASS is Craft evidence only — it is not Execution or Outcome for the piece being gated. Promoting a recurring failure into a new lint rule or banned-word tier is a separate harness-change item (floor E3/C3/O1, gates doctor, golden_check, pytest) and requires a golden corpus case proving the new check.

Constraints

All four checks run on every piece, each reported separately.

Four U's. Score each dimension 1-4. Thresholds: 12/16 for blog/SEO/articles, 10/16 for email/ads.

U	Question	Score
Unique	Can only WE write this? Original data, perspective, or experience?	1-4
Useful	Can the reader take action immediately?	1-4
Ultra-specific	Numbers, named tools, concrete examples?	1-4
Urgent	Is there a reason to engage today?	1-4

Banned words — instant reject (Tier 1): leverage, utilize, synergy, innovative, deep dive, circle back, touch base, moving forward, at the end of the day

AI slop (also reject): "In conclusion", "It's important to note", "In today's rapidly evolving", "This comprehensive guide", "Without further ado", "It's worth noting that"

Flag exact locations of every violation.

Voice patterns (programmatic — DO NOT skip). Binary clichés ("X-not-Y" / "It's not X, it's Y" / "isn't X — it's Y") read as LinkedIn slop and slip past subjective scoring. Run these regexes against the file with the Grep tool. Any match = FAIL.

Pattern	Catches
, not [a-z]	"X, not Y"
— not [a-z]	"X — not Y"
\bisn'?t [a-z][^.\n]+ — it'?s\b	"isn't X — it's Y"
\baren'?t [a-z][^.\n]+ — they'?re\b	"aren't X — they're Y"
\bIt'?s (a|the|an) [^.\n]+, not (a|the|an)\b	"It's a/the X, not a/the Y"
\bThat'?s (a|the|an) [^.\n]+, not (a|the|an)\b	"That's a/the X, not a/the Y"
\bIf you [a-z][^,.]+, [a-z]	"If you X, Y" rhetorical
\bHere'?s the thing\b	LinkedIn slop
\bI'?ll be honest\b	LinkedIn slop
\bLet that sink in\b	LinkedIn slop
\bHot take\b	LinkedIn slop

Skip matches inside HTML comments (<!-- ... -->) and code fences (```) — those are scorecard / metadata blocks.

Replacements that work: collapse to a single load-bearing claim; use parallel-positive contrast where both halves are positive ("Description is passive. State is something the agent can act on."); use a metaphor ("expensive webhook") rather than symmetrical reversal.

Project hook integration: if the project has .claude/hooks/voice-gate.py, that hook fires PostToolUse on Edit/Write and catches these patterns at draft time. The gate step still runs the regexes — belt and suspenders.

SEO lint (search content only). Apply only if the content targets search engines. Check against Algorithmic Authorship rules: conditions after main clause; instructions start with verbs; sentences under 20 words; bold the answer, not query terms; no links in first sentence of paragraphs.

Reporting. Report Four U's as a total out of 16 plus each dimension with its own score and reason; report banned words, AI slop, voice patterns, and SEO lint each as PASS / FAIL / SKIPPED with line numbers for every violation. On FAIL, name the specific fixes needed and offer to auto-fix and re-score.

Panel scoring is advisory and optional — only when the user asks for a second opinion or the artifact is high-stakes. Roles: audience reviewer (persona fit, plain-language clarity), channel reviewer (platform fit, format, norms), proof reviewer (source locations, attribution, unsupported claims), conversion reviewer (CTA clarity, objection handling), each scored 1-5 with concern and suggested fix. Label panel output as simulated review, never expert validation. Do not create fake credentials, endorsements, or named reviewers. Treat missing source locations as a gate issue, not a panel preference. Run panel scoring on drafts first; publishing still requires approval. Panel output cannot override a hard fail from banned words, source gaps, policy risk, or missing approval.

Context

Need	Load
Full Algorithmic Authorship rule set for SEO lint	knowledge/frameworks/content-copywriting/algorithmic-authorship.md
Per-format thresholds and gate minimums	harness/skill-contracts/ (the contract matching the piece's format)
Known loser patterns, and the write triggers for a repeat failure	memory/what-doesnt-work.md, memory/lessons.md, memory/MEMORY.md

Learning hook. Gate script runs log to data/learning/gate_runs.jsonl automatically. If the same diagnosis fails twice on one piece, append a lesson to memory/lessons.md before escalating (write trigger #3 in memory/MEMORY.md). Recurring failures across pieces get mined by /kai-retro and promoted into new gate checks with golden corpus cases.

Escalate when

• The same diagnosis fails twice on one piece — log the lesson, then hand the piece to a human with the specific failures listed.
• The content's format is unclear, so the correct Four U's threshold (12/16 vs 10/16) cannot be chosen.
• A violation sits in quoted source material or a client-supplied claim where rewriting changes the meaning.
• A check cannot run in this install — say so, report it SKIPPED, and never assume a pass.