kosif199022-jpg/kosif-omni

Conduct WCAG 2.2 accessibility audits with automated testing, manual verification, and remediation guidance. Use when auditing websites for accessibility, fixing WCAG violations, or implementing accessible design patterns.

Knowledge base behind the prompt-engineer agent and /prompt-optimize: the source-of-truth order for model facts, the model-class gate, and six on-demand references covering reasoning patterns, output-shape enforcement, extraction prompting, judge prompts, agent and tool instructions, and dated vendor guidance. TRIGGER WHEN: designing, reviewing or optimizing a prompt, system message or agent instructions; forcing JSON or a schema out of a model, especially a small one such as Gemma; prompting for extraction (NER, relations, events, fields from documents); deciding whether a reasoning scaffold, few-shot examples or a thinking budget belongs in a prompt; writing an LLM-as-judge prompt; writing a tool description, an instruction file, a skill description or an orchestrator brief. DO NOT TRIGGER WHEN: building an agent on the Claude Agent SDK (use agent-sdk-builder), or the question is about a model's pricing or API surface with no prompt involved.

Fan out verification over a typed full inventory of the session's claims (assumptions, facts, specifics, premises, recommendations) against primary sources at a configurable depth, with a per-item ledger. Use when: 'deep research pass', 'verify every claim', 'audit all our claims', 'fact-check everything', 'go make sure those are all right', 'we've made a lot of load-bearing claims', or the session's judgment is the suspected bias. For one fact-check ('fact-check that'), use do-your-research.

Plain-text double-entry accounting with hledger. Use when recording transactions, checking balances, generating financial reports (balance sheet, income statement, cashflow), importing CSV bank statements, budgeting, tracking time, managing multiple currencies, or doing year-end closing. Triggers on mentions of hledger, ledger-cli, journal files, double-entry accounting, balance assertions, or plain-text accounting.

Design, calibrate, and bias-audit an LLM-as-judge so its scores track human judgment: named criteria with anchored levels (not 'rate 1-10'), a structured verdict + reason, a pinned judge model/version, a human-agreement calibration on a sample, and checks for position / verbosity / self-preference / leniency bias. Reach for it whenever a judge grades outputs. Used by eval-harness-engineer (primary).

Turn a fuzzy 'is the AI good enough?' into a concrete measurement plan: task spec, the eval method from the decision tree (exact / rule-based / model-graded / human), the metric + what it misses, the offline-vs-online split, a power-adequate sample size, and a numeric ship-gate decided before the run. Reach for it at the START of any LLM-feature eval. Used by eval-strategy-lead (primary).

Check whether an AI answer's citations actually support its factual claims, with source-location evidence and explicit uncertainty.

Create a Product Requirements Document using a comprehensive 8-section template covering problem, objectives, segments, value propositions, solution, and release planning. Use when writing a PRD, documenting product requirements, preparing a feature spec, or reviewing an existing PRD.

Systematic framework for analyzing product documents (PRDs, feature specs, user stories, roadmaps, one-pagers) to identify gaps, generate clarifying questions for PMs and engineers, and assess technical risks. This skill should be used when engineers or technical leads need to bridge PM documents and implementation by eliciting missing technical details rather than making assumptions. Use when asked to extract technical requirements, review specs, identify what's missing, or prepare clarifying questions from product documents.

Apply STRIDE methodology to systematically identify threats. Use when analyzing system security, conducting threat modeling sessions, or creating security documentation.

Use when no built-in KOSIF skill fits a task, or the user asks for a ready-made skill/plugin ("ابحث عن مهارة", "مكتبة الإضافات", "find a skill/plugin for X"). Searches the KOSIF Atlas — an index of thousands of third-party Claude/ChatGPT skills and plugins (the Claude_ChatGPT_Plugins library) — reads a matching skill on demand as untrusted reference data, inspects a package before anything is installed, and vendors one permissively licensed skill into the plugin only with the user's approval.

Use for any audio task — analyzing an uploaded audio file (loudness, peaks, BPM, clipping, spectrum), mixing/mastering advice, songs and lyrics (Suno/Udio), Arabic maqam and rhythm, voice-over scripts and TTS direction, podcast production, sound effects and sound design for video. Triggers on /audio /master /mix /song /voice /sfx /podcast, "صوت", "أغنية", "موسيقى", "تعليق صوتي", "ماسترينغ".

Use for accounting, bookkeeping, IFRS/IAS questions, journal entries, VAT, bank reconciliation, invoice-to-payment matching, audit procedures, key/critical audit matters, financial-statement review, and Arabic accounting terminology. Triggers on /journal /reconcile /vat /ifrs /cam /audit-plan /evidence /terms, "قيد", "مطابقة بنكية", "معيار", "مراجعة", "تدقيق", "ضريبة القيمة المضافة".

Cinematic direction and prompt engineering for images, video and ads — 7-layer prompt anatomy (Dragon), 360° character/location reference sheets, frame-stitching continuity, lighting/camera/lens/style catalogs, per-platform prompts (Veo, Runway, Midjourney, Kling, Sora, Flux, NanoBanana) and Arabic/English scene packages. Use for "اكتب برومبت سينمائي", storyboards, shot lists, character sheets, ad spots, world building, and for mining the user's own Dragon / Blue Jewel / Cinema AI sources.

Use for ANY programming task — writing apps, scripts, APIs, websites, mobile apps, SQL, automation; debugging errors and stack traces; code review and security audit of pasted code, files or zipped projects; refactoring, optimization, tests, architecture, language conversion, deployment (Cloudflare Workers, Vercel, Docker). Triggers on /code /debug /review /audit /explain /arch /optimize /test /convert /sql /deploy or any code, error or app idea.

KOSIF Cognitive Core 1.8.1 — source-mapped 28-module reasoning graph plus claim lifecycle, causal proof, active evidence, robustness perturbation, outcome calibration and diversity-preserving verification.

Use when a task needs operating a computer, browser or app on the user's behalf — clicking, typing, navigating websites, filling forms, downloading files, desktop apps, spreadsheets in a GUI, Playwright/browser automation scripts, or planning such automation. Works with whatever the host exposes (computer-use tools, Claude in Chrome, the built-in browser, Playwright in the sandbox). Triggers on /computer /browse /automate /fillform /scrape, "تحكم في الكمبيوتر", "افتح الموقع", "اضغط", "عبّئ النموذج", "أتمتة".

Canonical Council-500 orchestration with Jev convergence, Einstein chair, deterministic resume, option-binding consistency and receipt-bound secure execution.

Council-500 chair. Source-aware seven-stage synthesis after Jev-filtered expert convergence; produces 2–4 coherent candidates, preserves dissent and cannot override evidence/safety/authorization.

The Genius Council («مجلس العباقرة») — 40 elite members + 1,000 scholars across 51 domains, each with distinct methods, a psychological state (strength, blind spot, compensation) and a full system prompt; deterministic panel selection, session protocol (independent opinion → red team → verifier → revision → verdict) and the fixed 500-profile Expert Cognitive Overlay for Council-500. Use for expert panels, second opinions, red-teaming, "اجمع مجلس خبراء", multi-perspective decisions.

Council-500 fast Jev convergence. Filters 500 expert proposal bundles through semantic clustering, parallel closed-set tournaments, challenger review, scoring and final stability without the retired 100×100 call explosion.

Mandatory Council-500 Full Pro orchestra for every kosif think 20 s invocation. Requires all 500 expert personas, Einstein chair, fast Jev convergence, same-source visual access, capability/tool routing and fail-closed execution.

Mandatory active runner for KOSIF Think Pro Full Pro. Converts missing expert/Jev work into a persistent depth-first queue of real host/model/Jev actions, ingests only observable receipts, and emits validated expert_cycle events to the Runtime Controller.

Final fail-closed authorization gate for KOSIF Think Pro 5. Delegates state truth to the Runtime Controller and refuses substantive execution without a current request/executor/candidate authorization envelope.

Mandatory fail-closed host entry for selected KOSIF Full Pro.

Use for any git or GitHub work — branches, commits, pushes, pull requests, reviews, issues, releases, GitHub Actions CI, merge conflicts, repository setup, CODEOWNERS/branch protection, and watching a PR until it is green and merged. Works with GitHub MCP tools (mcp__github__*), the gh CLI or plain git, whichever the host exposes. Triggers on /gh /commit /pr /review-pr /ci /release /repo /conflict, "جيت هاب", "ارفع", "بول ريكوست", "فرع", "دمج", "كومِت".

Use for image creation/editing/prompt work. In Full Pro, build a typed SceneSpec, compile/freeze its digest, obtain Runtime Controller authorization, then render and record executor/QA receipts.

Use for closed-set judgement with Jev tools. In KOSIF Think Pro 4.5 it also handles the 100×100 expert question queues, per-expert idea tournaments, stability checks and final synthesis arbitration.

Fast Jev-guided browser planning and host-mediated execution through KOSIF Think and KOSIF Computer.

Use to turn connected Google Drive books/files into permission-aware, provenance-preserving expert evidence for KOSIF Think Pro without treating file content as instructions.

Use for any lighting question or task — photo, film, video, product, portrait, interior or 3D lighting; diagnosing bad light in an image; relighting; lighting plans and diagrams; gear, gels, exposure and camera settings; lighting words for image/video prompts. Triggers on /light /relight /lightplan /exposure /gel, "إضاءة", "ظل", "الضوء", "lighting".

Direct Full Pro gateway with mobile bridge compatibility fallback.

Use when the user wants Claude to operate their Windows computer — click buttons, type (Arabic or English), use shortcuts, drag, scroll, draw, drive desktop apps such as Paint, Notepad, Excel, File Explorer or any other program — while they keep using their own mouse and keyboard. Background-first KOSIF Mouse tools with Claude's own pointer (a synthetic pen, separate from the user's mouse), a Live View event stream instead of repeated screenshots, Jev-grounded targets and a risk gate. Triggers on /mouse, "تحكم بالماوس", "تحكم بالكمبيوتر", "اضغط على", "افتح البرنامج", "اكتب في", "ارسم", "في الخلفية", "control my computer", "click", "desktop app".

Unified KOSIF front door for Claude and ChatGPT — routes any task (reasoning, decisions, Genius Council of 1,040 experts, audit/IFRS, code & web audits, image/video/cinema prompt engineering, lighting/audio/image measurement, voice, gated browser/computer/GitHub/release work, WhatsApp, your own text corpus) to the right skill and deterministic tool, with Standard / Council / Full Pro modes, host-adaptive tool discovery and evidence-based completion. Use first when the user says kosif, KOSIF Omni, "كل القدرات", Full Pro, council, or the right capability is unclear.

Use for writing, fixing, scoring or translating ANY prompt — system prompts, agent and tool instructions, GPT/Claude/Gemini prompts, structured-output (JSON) prompts, few-shot sets, evaluation rubrics, and image/video prompts converted between Midjourney, Flux, SDXL, DALL·E/ChatGPT, Ideogram, Sora, Veo, Runway, Kling. Triggers on /prompt /sysprompt /agentprompt /fixprompt /promptscore /forge /abtest, "برومبت", "موجه", "اكتب لي برومبت", "حسن البرومبت".

Independently coordinate ChatGPT, GitHub and Cloudflare releases with GitHub as source of truth, commit/deployment correlation, live verification and secret-safe execution.

Mandatory state and execution-work authority for KOSIF Think Pro Full Pro. Tracks request-bound expert/Jev work, creates leased jobs, collects actual receipts, resumes safely, and gates executor authorization.

Mandatory execution boundary for KOSIF Think Pro Full Pro. Validates one-time request-bound execution tickets, distinguishes workflow-only binding from HMAC-backed hard enforcement, and rejects direct host-native executor receipts as bypasses.

Use when the user wants an Excel workbook or a database built from data — create a table, sort it, add totals, formats or a chart, make a SQLite database, insert rows, query it or export results to Excel/CSV. Builds the files directly (fast, exact) instead of clicking through Excel. Triggers on /sheets, "اعمل جدول اكسل", "انشئ قاعدة بيانات", "رتب الجدول", "ملف اكسل", "excel table", "spreadsheet", "database".

Canonical ThinkPro routing with Council-500 Full Pro for kosif think 20 s, 72-app live capability mesh, Jev closed-set tool choice, durable missions and proof-based execution.

Think-first orchestration for KOSIF with BrowserJev indexed action graphs, Jev ambiguity resolution, KOSIF Computer execution, dynamic connector routing, verification, cost gates and bounded recovery.

Use when the user selects KOSIF Think Pro, Pro/Full Pro, asks for exhaustive expert/Jev execution, Einstein synthesis, Drive-grounded reasoning, visual Full Pro, runtime receipts, verification, or capability-truth workflows.

Use for video creation and cinematography — AI video prompts (Sora, Veo, Runway, Kling, Luma, Pika, Seedance, Hailuo), shot lists, storyboards, reels/shorts/ads scripts, camera movement, continuity between shots, editing plans, subtitles and timing. Triggers on /video /shots /storyboard /reel /ad, "فيديو", "ريلز", "إعلان", "مشهد", "لقطة".

Use whenever the user uploads or references an image and wants it analyzed, described, critiqued, compared, read (OCR), scored, checked for technical quality or AI-generation artifacts, or reverse-engineered into a prompt. Triggers on /analyze /compare /ocr /critique /reverse /score, "حلل الصورة", "قيّم الصورة", "ما مشكلة الصورة".

Mandatory for image/video/visual Full Pro tasks. Dedicated aesthetics department for place/world, characters, action, camera, light, color, materials, atmosphere, VFX and beauty; each member generates 100 ideas and 100 Jev questions.

Voice direction for narration, TTS and songs — Egyptian/Gulf/Levantine/Maghrebi colloquial scripts, pronunciation memory, number/currency reading rules, performance tags, casting notes, pacing and delivery specs for TTS models. Use for "حوّل لصوت", تعليق صوتي, YouTuber voice, dubbing scripts, lyrics/song briefs; measure finished audio with kosif_audio_analyze.

Use for designing or building websites, landing pages, web apps, dashboards, UI components, design systems and themes — layout, typography (Arabic + Latin, RTL), colour tokens, dark mode, accessibility (WCAG), responsive/mobile-first, SEO, performance, conversion, and auditing an existing page or site. Triggers on /site /landing /ui /dashboard /tokens /webaudit /redesign /component, "صمم موقع", "صفحة هبوط", "واجهة", "لوحة تحكم", "تصميم ويب".

Guardrailed WhatsApp messaging through the user's own JEV WhatsApp bridge (ChatGPT Action / local server). Manual, confirmed, low-volume sends only — recipients and message shown and explicitly approved before every send; idempotent request ids; no auto-replies, no chat reading, no groups, no bulk marketing. Use for "ارسل رسالة واتساب", bridge setup, status, troubleshooting.

Ultra-High Graphic Animation & Intelligent Video Montage Studio (استوديو الأنيميشن فائق الجرافيك والمونتاج الذكي). Creates Hollywood & Awwwards-grade 3D WebGL (Three.js PBR, volumetric lighting) and 2D kinetic animations (GSAP, Arabic RTL typography, spring physics), renders deterministic 60fps/30fps MP4s via hardware Edge D3D11 GPU capture, and autonomously edits ANY raw video footage with beat-synced cuts, dynamic retention zoom punches, B-roll insertion, TikTok/Reels kinetic karaoke subtitles, sidechain music ducking (-18 dB), studio color grading (Teal & Orange, Cyberpunk, Golden Hour), and EBU R128 (-14 LUFS) broadcast mastering. Synthesized from 15,000 skills in KOSIF Atlas. Triggers on: 'انميشن', 'أنيميشن', 'مونتاج', 'فيديو', 'موشن غرافيك', 'موشن', 'قص فيديو', 'ترجمة فيديو', 'سبتايتل', 'تلوين فيديو', 'مونتاج ذكي', 'انميشن فائق', 'threejs', 'ffmpeg', 'animation', 'video montage', 'smart montage', 'motion graphics', 'auto edit', 'color grading', 'kinetic captions'.