Kai hook bench

Build a ranked, provenance-tagged hook bank: sourced pains × formula families → score → gate → route. Hooks matching a known loser pattern are rejected at birth.

Scope boundaries: this skill produces the hook bank only. /kai-social batches full social posts, /kai-ad-campaign builds full ad campaigns, /kai-write writes single finished pieces, /kai-repurpose fans a pillar into derivatives — they consume this bank. Concept-level portfolio strategy (Persona × Desire × Angle bench, budgets, kill rules) belongs to knowledge/playbooks/combinatorial-creative-bench.md via /kai-ad-campaign; this skill fills the hook field of those bench rows.

Phase 0: Load Product Context

Check if MARKETING.md exists in the project root (same directory as CLAUDE.md, README.md, package.json).

If it exists: Read it — skip product discovery questions. It has the product name, ICP, value prop, monetization, brand voice, current channels, and competitive landscape.

If it does NOT exist: Auto-explore the codebase to create it in the project root (next to CLAUDE.md). Do NOT ask the user what the product is. Read CLAUDE.md, README.md, PROJECT.md, package.json, landing pages, and any project files. Search for email/ad/analytics config. Then create MARKETING.md using the template from /kai-email-system. Present draft to user for confirmation.

---

Phase 1: Load Inputs (pains, personas, winners, losers, proof)

Load all five before generating anything:

1. Personas — knowledge/personas/_persona-index.md. Pick 1-3 personas via its industry/pain tables (or the client's own from MARKETING.md). Note each persona's evidence status; hypothesis personas are fine for hook ideation but flag them in provenance.
2. Sourced pains — if workspace/offer-builder/pain-table.md exists (output of /kai-offer-builder Phase 1), use it: every row already has a source column. If it does NOT exist, mine pains with sources — load harness/references/audit-data-provenance.md, then:

   python -m scripts.audit.collect --url <business-url> --mode sales_external --workflow hook-bench --out workspace/hook-bench/data
   

   plus, where available: Reddit digests (hand off to /kai-reddit-listen if no profile exists in scripts/reddit_monitor/profiles/), user-provided call notes/reviews (cite file path), explicit WebSearch (URL + retrieval date), or python scripts/intel/brand_pulse.py <brand> --domain <domain>. Write workspace/hook-bench/pain-inputs.md with the same columns as the offer-builder pain table. No source, no row. Never invent "top pains from Reddit", review counts, or frequency percentages. Unsourceable pains go in workspace/hook-bench/_data-gaps.md.
3. Winners — knowledge/playbooks/what-works.md. If it has entries, note which hook styles won; new hooks should extend proven angles first. If empty (fresh install), record "no winners yet — all hooks are experiments" in provenance.
4. Anti-patterns (rejection list) — memory/what-doesnt-work.md + memory/lessons.md. Build workspace/hook-bench/rejection-rules.md: one row per banned pattern with the memory line that banned it. Seed rules currently in memory: the binary-contrast construction ("It's not X, it's Y" — reads as LinkedIn slop), and study percentages reframed as promises. Also load the voice-pattern regexes in harness/skills/kai-gate/SKILL.md step 3. Any candidate hook matching a rejection rule is discarded at generation time, logged to workspace/hook-bench/rejects.md with the rule that killed it — it never reaches ranking.
5. Proof inventory — proof-led hooks may only cite proof that exists on record. Use, in order: workspace/proof-library/ if present (owned by /kai-proof-builder — hand off there to create or refresh it, don't rebuild it here); case studies already in workspace/ (produce new ones via /kai-case-study); data/content_log.json 30-day winners; collector output from step 2. Write workspace/hook-bench/_proof-inventory.md: claim, source path/URL, permission status. An empty inventory is a finding (log it in _data-gaps.md), not a license to fabricate testimonials or numbers.

Log every source in workspace/hook-bench/_sources.md (URL/path, method, retrieval date).

Phase 2: Hook Formula Library

Load the three mechanics sources and extract the mechanics they state — no vague gestures:

• knowledge/people/alex-hormozi-knowledge.md — the hook mechanics live in "$100M Leads": the Hook–Retain–Reward content formula (hook = bold claim, counterintuitive statement, or direct problem identification; retain = unresolved questions, storytelling, cliffhangers), the paid-ads rule (hook in first 3 seconds, appeal to identity or location, then Problem-Agitation-Solution body), and the lead magnet naming formula ([Number] + [Adjective] + [Target Audience] + [Desired Outcome] + [Timeframe]). From the same doc's other sections: the Value Equation (Dream Outcome × Perceived Likelihood of Achievement ÷ Time Delay × Effort & Sacrifice), the five content pillar categories (business frameworks/own IP, common mistakes, counterintuitive takes, behind-the-scenes, Q&A and consulting replays), the CLOSER "S" step ("Sell the Vacation, Not the Plane Flight"), and "People don't buy products. They buy certainty of outcomes."
• knowledge/playbooks/ad-creative-best-practices.md — the 3-Second Rule, the hook_type taxonomy (problem, proof, mechanism, contrast, offer, story, pattern interrupt), and the PAS/AIDA/BAB copy formulas.
• knowledge/playbooks/combinatorial-creative-bench.md — angle types (loss-aversion, social proof, mechanism, contrast, time math) and the rule that a hook is an execution variable inside a P.D.A. concept.

Organize into these seven formula families (each is supported by the sources above — cite the exact section per family in the library file):

Family	Mechanism	Grounding
pain-callout	Name the specific sourced pain in the reader's language	Hormozi "direct problem identification"; PAS Pain step; persona-index hook lines
identity-callout	Name who this is for so the persona is obvious in 3 seconds	Hormozi paid-ads hook rule ("appeal to identity or location"); bench persona-clarity score
dream-outcome	Lead with the vivid destination, status-aware, not the feature	Value Equation numerator; CLOSER "sell the vacation"; BAB After step
proof-led	Open with a real result, demo, or named example	"Certainty of outcomes"; hook_type: proof; TikTok proof-first hook (knowledge/channels/tiktok-algorithm.md)
curiosity-gap	Bold or counterintuitive claim that opens an unresolved question	Hormozi content pillar "counterintuitive takes" + Retain mechanics; AIDA Attention
common-mistake	Contrast what the ICP does wrong with what works	Hormozi content pillar "common mistakes"; hook_type: contrast; bench contrast angle
speed-effort	Collapse the Value Equation denominator: outcome + timeframe + low effort	Lead magnet naming formula; MAGIC Interval; Time Delay / Effort & Sacrifice variables

Write workspace/hook-bench/formula-library.md: per family — mechanism, 2-3 fill-in patterns, source citation, and known failure mode (e.g. common-mistake degenerates into the banned binary-contrast cliché; curiosity-gap degenerates into clickbait that the body can't reward; speed-effort degenerates into unsubstantiated timeframe claims, which fail the compliance pass).

Phase 3: Generation Matrix

Generate N hooks per pain × family cell (default N=2; take the top 5 pains by frequency signal → up to 70 candidates). For each candidate record: hook_id (HB-{pain#}-{family}-{nn}), pain row ref, persona, family, draft text, proof ref (or none), target platforms.

Rules at generation time:

• Reject at birth: run every candidate against rejection-rules.md and the kai-gate voice regexes before it enters the matrix. Log kills to rejects.md.
• Quantitative claims ("booked 40 calls", "took 6 weeks") only from _proof-inventory.md or Phase 1 sourced data — otherwise rewrite qualitative or drop.
• Platform ceilings (verified against the cited files — respect the tightest ceiling of the hook's target platforms):

Surface	Constraint	Source
X standard post	280 chars total	knowledge/channels/twitter-x.md
LinkedIn post	hook lands in first ~210 chars (mobile "...see more" fold)	knowledge/channels/linkedin-organic.md
Meta ads	primary text 125 chars recommended; headline 40 chars	knowledge/playbooks/ad-creative-best-practices.md
TikTok	visual hook in first 0.7-2s, must work on mute; ad text 100 chars	knowledge/channels/tiktok-algorithm.md; knowledge/playbooks/ad-creative-best-practices.md
Google Search ads	headline 30 chars	knowledge/playbooks/ad-creative-best-practices.md

Write the full matrix to workspace/hook-bench/generation-matrix.md.

Phase 4: Ranking + Gate

4a. Rubric. Score every surviving candidate 0-2 per dimension (0 = fails, 1 = weak, 2 = strong):

Dimension	2 means
Clarity	Understood in one read, works out of context, no setup needed
Specificity	Named persona/situation/number — could not be about any product
Curiosity	Opens a question the reader needs the next line to close
Proof-backing	Claim is backed by an _proof-inventory.md row or sourced pain quote

Hard rule: a proof-led hook with proof-backing 0 is cut or rewritten into another family — it may not ship as proof-led. Justify each score in a phrase, citing the pain row or proof ref. Rank by total (/8); ties break by winner-adjacency (Phase 1 step 3).

4b. Gate the top set. Take the top ~25-30 hooks into hook-bank.md (Phase 5 format), then run:

python scripts/quality_gates/four_us_score.py --file workspace/hook-bench/hook-bank.md    # 10/16 — ad/hook threshold
python scripts/quality_gates/banned_word_check.py --file workspace/hook-bench/hook-bank.md

Max 2 retry cycles; fix only the named failing dimension or word, never rewrite the bank (see memory/lessons.md). After 2 failures, escalate to a human with the diagnosis and log it via python scripts/self_improvement/lesson_capture.py add. Record scores, retries, and rubric table in workspace/hook-bench/_gate-report.md.

Phase 5: Hook Bank Output

Write workspace/hook-bench/hook-bank.md, ranked. Each entry:

### #3 · HB-2-proof-led-01 (score 7/8)
- short  (overlay / one-liner, ≤ 60 chars — fits every ceiling above): "..."
- mid    (caption / primary text, ≤ 125 chars — Meta primary-text ceiling): "..."
- long   (opener, 1-3 sentences, ≤ 210 chars before the fold — LinkedIn/long-caption first lines): "..."
- provenance: pain #2 (source URL/path, date) · family: proof-led · proof: _proof-inventory.md row 4 · persona: Admin Martyr (directional)
- routing: /kai-ad-campaign (Meta TOF) · /kai-social (TikTok overlay)

The three lengths carry the same promise — same pain, same claim, no escalation of the claim as length grows.

Routing notes (append a routing table to the bank):

Consumer	Use	They own
/kai-social	short + mid as post/overlay openers	full post body, hashtags, schedule; contract harness/skill-contracts/social-post.yaml
/kai-ad-campaign	short as ad headline seed, mid as primary text seed, hook_id fills the bench row's hook field	platform policy check (per-platform table in .claude/rules/architecture-and-memory.md + harness/references/ad-write-guardrails.md), funnel mapping, upload
/kai-write	long as first-line/opener for single pieces (script, email, article)	framework, contract, full draft

Approval doctrine: the bank is an internal asset — nothing here posts, uploads, or mutates a live channel. Downstream skills carry their own gates and human approval before anything ships.

Feedback loop: when downstream 30-day checks grade a hook-led piece, winners feed knowledge/playbooks/what-works.md automatically; graded underperformers get their hook family/pattern diagnosed into memory/what-doesnt-work.md via /kai-retro, which tightens rejection-rules.md on the next run.

Output Tree

workspace/hook-bench/
├── _sources.md              # every source: URL/path, method, retrieval date
├── _data-gaps.md            # unsourceable pains, empty proof inventory, unknown metrics
├── _proof-inventory.md      # sourced proof rows (or pointer to workspace/proof-library/)
├── data/                    # scripts.audit.collect output (kai-data.json), if mined here
├── pain-inputs.md           # Phase 1 pains (only if workspace/offer-builder/pain-table.md absent)
├── rejection-rules.md       # anti-patterns from memory/ + kai-gate regexes
├── rejects.md               # hooks killed at birth, with the rule that killed each
├── formula-library.md       # Phase 2 — 7 families, patterns, citations, failure modes
├── generation-matrix.md     # Phase 3 — full pain × family candidate set
├── _gate-report.md          # Phase 4 — rubric scores, Four U's / banned-word results, retries
└── hook-bank.md             # Phase 5 — ranked bank, 3 lengths, provenance, routing

Hand-offs (do not re-specify these jobs)

• Pains not yet mined → /kai-offer-builder Phase 1 (its pain table is this skill's preferred input)
• Proof library missing or stale → /kai-proof-builder · customer-story proof → /kai-case-study
• Ongoing pain listening → /kai-reddit-listen
• Hooks into posts → /kai-social · into ads → /kai-ad-campaign · into single pieces → /kai-write (brief first via /kai-brief)
• Independent re-gate of the bank → /kai-gate
• Diagnosing underperforming hooks after 30-day grades → /kai-retro