Ah Sheli Gibor

אח שלי גיבור — the affectionate Israeli address. The comprehensive Hebrew production suite. 25+ output types · 15 variation modes · 6 voice personas · 124-source AI catalog · six-stage validation · 4-axis Hebrew-labeled rubric. Production-grade.

What you get

A Claude Skill that produces professional-grade Hebrew across every format and register you actually need — books, articles, speeches, pitches, research papers, executive reports, course materials, talking cards, teleprompter scripts, technical documentation, and more. With the right voice, the right register, validated grammar, and source-aware generation.

This is not a translator. It's a Hebrew producer that comprehends source content and reconstructs it in the target format, persona, and variation mode the situation needs.

Why this exists

Israeli Hebrew is fragmented. A pitch to investors needs different vocabulary than a memorial speech. A medical clinical note needs different grammar than a haredi-tech LinkedIn post. A cybersecurity incident report needs different cadence than a poetry chapbook. Generic translation tools flatten all of this into the same beige Hebrew.

The Ah Sheli Gibor suite gives you 25+ format specs × 15 register modes × 6 voice personas × 124 AI sources. Every combination produces Hebrew that a 2026 Israeli professional in that specific context would actually write.

What makes it trustworthy

• Anti-hallucination guards — every term traces to corpus (2025+ source-dated) or a documented persona signature
• Provenance tracking — corpus entries carry URL + date; sources cited
• Academy of Hebrew Language alignment — formal rulings respected
• 2025-only source rule — no dated 2022–2023 buzzwords
• Deterministic legal output — contracts produce reproducibly (temperature=0)
• 6-stage validation — every output checked across grammar, jargon currency, persona, phrasing, anti-patterns, rubric
• 4-axis Hebrew-labeled rubric — רלוונטיות / קוהרנטיות / עקביות / רהיטות, each 1-10
• Native-speaker authenticity test as the final pass

Installation

One-line install (recommended)

npx ah-sheli-gibor

Installs from the npm registry to ~/.claude/skills/ah-sheli-gibor/. Restart Claude Code afterward.

Alternative — from GitHub source

npx github:raifalex/Ah-Sheli-Gibor

Global install

npm install -g ah-sheli-gibor
ah-sheli-gibor

Installer options

npx ah-sheli-gibor                  # default install (~/.claude/skills/)
npx ah-sheli-gibor --update         # git pull latest
npx ah-sheli-gibor --uninstall      # remove
npx ah-sheli-gibor --target <path>  # custom path
npx ah-sheli-gibor --dry-run        # preview without installing
npx ah-sheli-gibor --help

Requires Node ≥14 and git.

Optional Hebrew NLP toolkit

For the automated validator + 14 specialized-model subcommands:

cd ~/.claude/skills/ah-sheli-gibor
pip install -r scripts/requirements.txt   # transformers + torch + sentencepiece

Supported agents

claude-code · cursor · github-copilot · windsurf · opencode · codex · gemini-cli

Quick examples

Example 1 — Founder LinkedIn announcement

Input:

"Write a LinkedIn announcement for closing a $14M Series A. Voice = Yoel. Add details: 22 enterprise customers, $2.1M ARR, 142% NRR."

Output (Yoel voice):

רגע נדיר.

אחרי 18 חודשים של בנייה, שלושה rewrites של ה-architecture, ופיבוט אחד שכמעט הרג אותנו — סגרנו היום Series A של 14 מיליון דולר.

מה למדנו: רוב הצוותים שמנסים להטמיע AI agents מתקשעים על אותה בעיה — context handoff בין agents. כולם פתרו את זה לרוחב; אנחנו פתרנו לעומק. זה bet ששילם.

22 לקוחות enterprise. $2.1M ARR. NRR של 142%. צוות של 9 שבונה כמו 30.

מגייסים founding engineers עכשיו. אם זה מדבר אליך — דבר איתי.

Example 2 — Executive board memo

Input:

"Board memo recommending we raise the GTM budget by $4.2M for Q3-Q4. Persona = Yoel. Format = report-executive."

Output (excerpt, Hebrew-formal):

לכבוד: דירקטוריון SecureFlow מאת: דני שריג, מנכ"ל תאריך: 22/05/2026 נושא: המלצה לאישור הרחבת תקציב GTM ב-4.2 מיליון דולר

תקציר מנהלים

מאז סבב ה-Series B צמחנו פי 2.4 ב-ARR. ה-pipeline ב-Enterprise תפח ב-340%. בקצב הנוכחי אנחנו מתקרבים לחסם עקב חוסר ב-AE capacity. מבקשים אישור להגדלת תקציב ה-GTM ב-4.2 מיליון דולר.

המלצה: לאשר.

[... full memo continues with Background / Analysis / Risks / Next Steps ...]

Example 3 — Cybersecurity panel talking cards

Input:

"Talking cards for a TechGym infosec panel on AI agent security. Persona = Dana. Variation = cybersecurity."

Output (excerpt):

CARD 1.1 ▶ 00:00 · OPENING CLAIM

השאלה היא לא איך — השאלה היא מתי.

ה-Shared Responsibility Model שלכם הניח אדם. וההנחה הזאת שגויה היום בכל סביבת ענן שמריצה AI agent.

זה לא קר. זה מדויק.

◼ [BEAT] ◼

The 25+ output types

| Category | Output types | |---|---| | Stage / spoken | rewrite · pitch · speech · talking-cards · teleprompter | | Books | book-chapter · book-proposal · manuscript-edit | | Articles | article-feature · article-op-ed · article-news · article-profile | | Educational | course-material (syllabus / lesson / handout / assessment / module / reading-guide) | | Reports | report-executive · report-business · report-whitepaper · report-incident | | Academic | research-paper (IMRAD) · research-proposal · thesis-chapter | | Business | business-plan · rfp-response (Israeli michraz) · case-study | | Communications | comms (press-release / formal-email / memo) | | Technical | tech-doc (README / API / runbook / ADR / migration / troubleshooting) · product-spec (PRD) |

Each output type has its own structural template, validation gates, and persona pairings. Full specs in output_types/.

The 15 variation modes + 4 sub-filters

Tech sub-domains (8)

| Mode | When to use | |---|---| | tech-general (default) | General Israeli tech writing — fallback | | software-engineering | Developer / DevOps / SRE register | | cybersecurity | Israeli infosec community (SOC / IR / DFIR / INCD) | | product-management | PM voice — roadmap / OKRs / KPIs / NPS | | defense-aerospace | Israeli defense (Elbit / MAFAT / Rafael / IAI) | | ai-ml-research | Academic + industrial ML | | startup-fundraising | Founder ↔ investor | | gen-z-creator | TikTok / Instagram / podcast |

Domain specialized (3)

| Mode | When to use | |---|---| | legal-technical | Contracts, ToS, IP — deterministic output | | medical | Clinical, patient comms, drug labels | | biblical-rabbinic | Religious, ceremonial, Talmudic |

Voice / style (4)

| Mode | When to use | |---|---| | gender-emotional | Personal narrative, vulnerable, memorial | | slang-cultural | Casual + cultural-explanation layer | | bilingual | Hebrew + English side-by-side | | creative-lyrics | Poetry, lyrics, experimental |

Community sub-filters (combine with any base mode)

| Sub-filter | When to use | |---|---| | arabic-hebrew-bilingual | Arab-Israeli, Druze, Circassian tech professionals | | haredi-tech | Bnei Brak / Beit Shemesh tech (9,700+ workers, 6,900 women) | | academic-formal | Israeli university research register | | diaspora-israeli | Bay Area / NYC / Berlin Israeli expats |

Full spec: references/hebrew_variations.md.

The 6 personas

| Persona | M/F | Archetype | Signature | |---|---|---|---| | יואל ״יו־יו״ שריג | M | tech-founder | Series-A confidence, dense English code-switching | | שירה לב | F | literary speechwriter | Classical-modern Hebrew, recurring images | | גלעד אש | M | comedian | Deadpan, slang-fluent, Hebrew setup + English punch | | דנה אלמוג | F | TV panelist-pundit | Debate-trained soundbites, rhetorical-question framing | | איתמר חוזה | M | veteran journalist | Patient long-form, classical Hebrew, deferred conclusions | | נועה אופק | F | contemporary creator | Intimate, vulnerable, fluid Hebrew-English |

Pick by name ("voice = יואל" / "speak as שירה") or let the skill auto-select based on output type + goal + variation mode.

Full persona profiles: personas/.

The interview (STEP 0)

When invoked, the skill asks (at most 3 questions, combining related ones):

| # | Variable | Example | |---|---|---| | 1 | Output type | rewrite / pitch / speech / book-chapter / research-paper / etc. | | 2 | Use context | who's the audience, what platform | | 3 | Purpose | inform / persuade / entertain / sell / mobilize / celebrate / mourn | | 4 | Mood | confident / warm / urgent / measured / playful / serious / vulnerable | | 5 | Goal — what you want to achieve | "convince investors" / "win a panel debate" / "land emotional impact" / "ceremonial address" / "sign a contract" / "clinical documentation" / "publish a chapter" / "submit a grant" | | 6 | Variation mode | (1 of 15) + optional sub-filter | | 7 | Persona | יואל / שירה / גלעד / דנה / איתמר / נועה / auto |

If 5/7 are clear from context, the skill proceeds — naming its inferences in one line.

The 4-axis output rubric

Every output is scored (1–10 each):

| Hebrew | English | Definition | |---|---|---| | רלוונטיות | Relevance | Coverage and priority of key content from the source | | קוהרנטיות | Coherence | Collective quality of all sentences; logical flow and arc | | עקביות | Consistency | Factual fidelity between output and source; no hallucination | | רהיטות | Fluency | Grammar, spelling, punctuation, word choice, sentence structure, persona fidelity |

Per-output-type pass thresholds:

| Output type | Priority axis | Threshold | |---|---|---| | Rewrite | רהיטות | All ≥ 7 | | Pitch | קוהרנטיות | קוהרנטיות ≥ 8 | | Speech | קוהרנטיות + רהיטות | both ≥ 8 | | Talking cards | עקביות | עקביות ≥ 9 (debate-grade) | | Teleprompter | רהיטות | רהיטות ≥ 9 (reading-speed) | | Research paper | עקביות | עקביות ≥ 9 (factual rigor) | | Legal contract | עקביות + רהיטות | both ≥ 9 (deterministic + precise) | | Incident report | עקביות | עקביות ≥ 9 (audit-grade) |

Goal further weights axes. Full spec: references/output_evaluation_rubric.md.

Six-stage validation

Every output passes through:

| Stage | What it checks | Tools | |---|---|---| | 5a Regex grammar | Pattern-based errors (Categories A/D/E/G/H/K/L) | hebrew_validate.py --no-model | | 5b Model grammar | Agreement, smikhut, binyan, gender/number | DictaBERT-parse | | 5c Jargon currency | Every term traces to corpus (2025+) or persona | corpus/jargon.json + Academy | | 5d Persona consistency | Voice fingerprint across paragraphs | personas/*.md | | 5e Phrasing / idiomaticity | Word order, idioms, register, code-switching density | references/phrasing_checker.md | | 5f Anti-patterns + authenticity | 12-category error catalog + native-speaker test | references/common_errors_catalog.md | | 5g 4-axis rubric | רלוונטיות / קוהרנטיות / עקביות / רהיטות (1-10 each) | LLM-graded with goal/output-type weighting |

Below threshold on any priority axis = automatic rewrite of the failing section.

The 124-entry source catalog

Consolidated from:

• Daniel Rosehill — Hebrew-AI-Models (CC BY 4.0)
• Daniel Rosehill — Hebrew-LLMs (CC BY 4.0)
• NVIDIA Developer Blog — TensorRT-LLM for Hebrew
• NNLP-IL Hebrew-Resources
• Academy of the Hebrew Language

By category:

| Category | Count | Top picks | |---|---|---| | LLMs (generation/reasoning/instruction) | 30 | DictaLM-3.0-24B-Thinking, Hebrew-Mixtral-8x22B | | ASR | 15 | ivrit-ai whisper-large-v3-turbo-ct2 | | TTS | 9 | SIMS-7B + phonikud, HebTTS | | BERT foundation | 11 | DictaBERT, NeoDictaBERT-bilingual, HeRo | | NER | 3 | dictabert-ner, heBERT_NER | | Sentiment / emotion | 4 | heBERT_sentiment, hebEMO (8 categories) | | Morphology / parsing | 7 | dictabert-parse, dictabert-large-parse | | Diacritization | 3 | Dicta Nakdan API | | Translation | 3 | Helsinki-NLP opus-mt, DeepL API | | Summarization | 4 | het5_summarization | | Domain-specialized | 6 | Legal-heBERT, hebrew_medical_ner_v5, BEREL_3.0 | | Embeddings | 3 | sentence-transformers-alephbert | | OCR / vision | 3 | testing-trOCR-hebrew-handwritten | | Speech foundation | 4 | mhubert-base-25hz, StresSLM, PAST | | Authoritative references | 4 | Academy of Hebrew Language, Morfix, Rav-Milim, MILA | | Runtime optimization | 1 | NVIDIA TensorRT-LLM (DictaLM-2.0) |

Full structured catalog: sources/hebrew_ai_models.json + sources/source_index.md + sources/source_selection.md.

The hebrew_toolkit.py CLI

14 subcommands invoking specialized Hebrew models on demand. Lazy-loads per subcommand.

python scripts/hebrew_toolkit.py morph ״ההנחה הזאת...״         # DictaBERT-morph
python scripts/hebrew_toolkit.py parse "..."                    # DictaBERT-parse → dep tree
python scripts/hebrew_toolkit.py ner "..."                      # DictaBERT-NER
python scripts/hebrew_toolkit.py sentiment "..."                # heBERT_sentiment
python scripts/hebrew_toolkit.py emotion "..."                  # hebEMO → 8 emotion scores
python scripts/hebrew_toolkit.py legal "..."                    # Legal-heBERT embeddings
python scripts/hebrew_toolkit.py medical "..."                  # medical NER
python scripts/hebrew_toolkit.py metaphor "..."                 # hebert-metaphor
python scripts/hebrew_toolkit.py offensive "..."                # offensive-detection
python scripts/hebrew_toolkit.py nakdan "..."                   # Dicta Nakdan diacritization
python scripts/hebrew_toolkit.py translate "..." --to en        # Helsinki-NLP MT
python scripts/hebrew_toolkit.py summarize @article.txt         # het5_summarization
python scripts/hebrew_toolkit.py recommend --task X --variation Y
python scripts/hebrew_toolkit.py rubric output.txt source.txt   # 4-axis template

Output: structured JSON. Input: literal Hebrew text or @filepath.

Production deployment

For high-volume Hebrew LLM serving: DictaLM-2.0-Instruct + NVIDIA TensorRT-LLM + Triton Inference Server. Near-constant latency at 16+ concurrent requests on a single A100.

Cost crossover from Claude API to self-hosted: ~5–10M Hebrew tokens/month.

Full recipe: references/nvidia_tensorrt_optimization.md.

When NOT to use this skill

• General Hebrew translation → use DeepL or DictaLM directly
• Hebrew RTL CSS / web layout → use hebrew-rtl-best-practices
• Hebrew PDF / DOCX / PPTX physical generation → use hebrew-document-generator
• Niqud / vowelization only → use Dicta Nakdan directly
• Non-Hebrew content

File structure

ah-sheli-gibor/
├── SKILL.md                            # Operating instructions (the protocol)
├── metadata.json                       # Bilingual skill metadata
├── package.json                        # npm package + bin
├── README.md                           # This file
├── CONTRIBUTING.md                     # How to add corpus entries
├── LICENSE                             # MIT
│
├── corpus/jargon.json                  # Vocabulary corpus with provenance
│
├── personas/                           # The 6 voices
│   ├── yoel-yoyo-sarig.md              ├── shira-lev.md           ├── gilad-esh.md
│   ├── dana-almog.md                   ├── itamar-hoze.md         └── noa-ofek.md
│
├── output_types/                       # 25+ output types (NEW v0.5.0 expanded)
│   ├── pitch.md   speech.md   talking_cards.md   teleprompter.md
│   ├── book-chapter.md   book-proposal.md   manuscript-edit.md
│   ├── article-feature.md   article-op-ed.md   article-news.md   article-profile.md
│   ├── course-material.md
│   ├── report-executive.md   report-business.md   report-whitepaper.md   report-incident.md
│   ├── research-paper.md   research-proposal.md   thesis-chapter.md
│   ├── business-plan.md   rfp-response.md   case-study.md
│   ├── comms.md   tech-doc.md   product-spec.md
│
├── sources/                            # 124-entry AI catalog (v0.4.0)
│   ├── hebrew_ai_models.json   hebrew_llms.json   source_index.md   source_selection.md
│
├── references/
│   ├── grammar_layer.md                # Binyan / gender / smikhut / preposition rules
│   ├── grammar_validation_tools.md     # DictaBERT / Hspell / Nakdan toolchain
│   ├── common_errors_catalog.md        # 12-category error catalog (A-L)
│   ├── anti_patterns.md                # Bad-output table
│   ├── phrasing_checker.md             # Idiomaticity / naturalness layer
│   ├── hebrew_variations.md            # 15 variation modes (NEW v0.5.0 expanded)
│   ├── output_evaluation_rubric.md     # 4-axis Hebrew-labeled scoring
│   └── nvidia_tensorrt_optimization.md # Production deployment
│
├── research/contemporary_voices_2026.md
│
├── scripts/
│   ├── hebrew_validate.py              # Fast regex + DictaBERT validator
│   ├── hebrew_toolkit.py               # Unified Hebrew NLP CLI (14 subcommands)
│   └── requirements.txt
│
├── bin/install.js                      # npx installer
├── tests/                              # Test cases + results
└── examples/rewrites/                  # Before/after examples

Versioning

• v0.1.0 — scaffold + 30 seed corpus + 5 tests + rewrite-only
• v0.1.1 — npx installer
• v0.2.0 — 6 personas + 4 output types + interview + initial validation
• v0.3.0 — 6-stage validation + phrasing checker + grammar tools + DictaBERT validator
• v0.4.0 — 124-source catalog + 8 variation modes + STEP 4.5 source selection + STEP 5g 4-axis rubric + user-goal interview + hebrew_toolkit.py (14 subcommands)
• v0.5.0 (current) — comprehensive scope expansion + rebrand as the Hebrew production suite: 25+ output types (added books / articles / courses / reports / research / business / communications / technical docs) + 15 variation modes (added software-engineering / cybersecurity / product-management / defense-aerospace / ai-ml-research / startup-fundraising / gen-z-creator) + 4 community sub-filters (arabic-hebrew-bilingual / haredi-tech / academic-formal / diaspora-israeli)
• v0.6.0 (planned) — audio rehearsal loop (TTS + ASR feedback); visual deliverable pipeline (markdown → PDF / Gamma decks); corpus expansion to 200+ 2025–2026 web-sourced entries
• v0.7.0 (planned) — educational mode; strict-corpus mode; self-improvement feedback loop; custom persona from user samples
• v1.0.0 (planned) — 300+ corpus entries, additional personas, CI / GitHub Action integration

License

MIT. See LICENSE. Underlying model licenses (DictaBERT CC BY 4.0; Hspell AGPL-3.0; Gemma / Llama / Mistral per their respective licenses) — verify before commercial use.

Contributing

See CONTRIBUTING.md. Pull requests welcome — especially for:

• 2025–2026 corpus entries from Israeli sources
• New variation-mode entries
• New output-type specifications
• Test cases for personas × variation modes × output types
• Custom persona profiles built from public Israeli voice samples

Built with the help of Claude Opus 4.7 (1M context). Catalog sources: Daniel Rosehill, NVIDIA Developer Blog, dicta-il, ivrit-ai, avichr, yam-peleg, onlplab, imvladikon, slprl, HeNLP, Helsinki-NLP, Norod78, Slasky, thewh1teagle, Intel, NNLP-IL.