digitizers/agent-skills

Operate the Canva MCP for branded design work — social posts, presentations, marketing collateral, one-pagers. Use when the user wants to create or edit a Canva design, fill a brand template, resize for another format, or export a design ("Canva", "פוסט לרשתות", "מצגת", "brand template", "עצב לי", "make a deck in Canva"). Not for AI image generation or upscaling (magnific-studio), not for UI/product mockups or design-to-code (figma-studio), and not for building slide files locally (the pptx skill).

Drive a pull request to convergence through the Codex AI reviewer — build → local Round 0 pre-review → PR → @codex review → verify each finding against HEAD → fix the real ones with regression tests → re-trigger until clean → human reviews last. Use when a PR is open or just pushed and should be reviewed, when a branch is built and tested and a PR is about to be opened (Round 0 local pre-review), when the user mentions "codex", "@codex review", "the review loop", "ultrareview", or asks to iterate a PR to green.

Use this skill IMMEDIATELY whenever the message contains "fable-mode", "fable mode", "פייבל", or "מצב פייבל" — including as a prefix before a task (e.g. "fable-mode: review this...") — no matter what the task itself is. Also use it proactively, unprompted, in two situations: (1) the user wants a fact double-checked or verified against a live source instead of answered from memory — "double-check", "can you verify", "is this still true", "I don't trust my memory", current versions/limits/pricing/capabilities of tools, APIs, or services; (2) the task is costly if wrong — production changes, migrations, deletions, security, money, architecture commitments, debugging, or research synthesis. Enforces strict planning, verification, and calibrated claims; stays active for the rest of the conversation until "fable-mode off" / "כבה מצב פייבל". Do NOT use for casual questions, trivial edits, creative writing, or when "fable" appears only as an ordinary word (writing a fable, defining it).

Route design work through the Figma MCP — client-site mockups, UI screens, design-to-code, code-to-design, FigJam diagrams. Use when the user wants to mock up a page or screen, implement a Figma design as code, push a page into Figma, diagram a flow, or shares a figma.com URL ("Figma", "FigJam", "mockup", "מוקאפ", "design-to-code"). Not for branded social posts or presentations (canva-studio) and not for AI image generation or upscaling (magnific-studio).

Audit and cut GitHub Actions minutes consumption in a repo — measure real usage first, hunt REDUNDANT work before optimizing existing work (does this job need to run at all? at this frequency? only then tune it), detect double-builds against deploy platforms (Vercel/Netlify/Cloudflare Pages/Cloudways), then propose exact diffs grouped by delete/reduce-frequency/optimize. Use whenever the user's intent is to SPEND LESS on CI — reduce GitHub Actions minutes, audit CI cost or billing, ask why Actions usage is high, ask what CI work can be cut or is redundant, request a CI-efficiency review of .github/workflows/, or mention "gha-optimizer". Read-only by default — never edits workflows without explicit approval. The dividing line is intent, not subject matter: this skill answers "what is our CI costing us, and what can we delete?" It is the wrong skill for "make this CI work" — writing a new workflow, or fixing a red/failing/misconfigured job — where the user wants MORE working CI, not less CI.

Use when the current session's work must continue in a fresh agent session — the context window is nearly full or about to be compacted, the user is ending a work session and will resume later, the work is being handed to another agent, machine, or teammate, or the user asks for a "handoff", "handoff document", "session summary to continue from", or to "prepare this for the next session". Also fires when a hook or system reminder reports high context usage and asks for a handoff. Not for writing project documentation for humans (README, docs, changelogs, reports) — only for transferring this session's state to a future agent session.

Magnific Studio — generate AI images and upscale via the Freepik / Magnific REST API. Use when the user wants to generate an AI image, hero visual, banner background, illustration, or texture from a text prompt, or to upscale/enhance an existing image. Triggers on "Magnific", "Freepik", "generate an image", "AI background", "AI hero visual", "text-to-image", "upscale this image". Not for branded social posts or presentations (canva-studio) and not for UI mockups (figma-studio). Reads the API key from ~/.claude/freepik.env (FREEPIK_API_KEY) so it works across all sessions.

Default to a branch + pull request for every change — code, docs, config, content, and marketing copy alike — instead of committing to main. Branch → commit → local pre-review (Round 0) → push → PR → review → merge → return to main; direct-to-main only when the human explicitly says so. Use when starting any change in a git + GitHub repo, when deciding whether something can skip a PR, when the user mentions branching / PR workflow, and whenever a request implies landing a change — "commit it", "ship it", "push it", "get it into main", "publish it", "update the copy / the page / the README" — however small the edit sounds. Pairs with codex-review-loop, which pre-reviews the built branch (Round 0) and drives the opened PR to green.

Safely run a one-off write, backfill, or data-mutating script against a PRODUCTION database — pull the connection from the platform, dry-run, get explicit human authorization, execute, verify, clean up. Use before running any script that inserts/updates/deletes prod data (generating codes, backfills, one-off fixes, seed data), when the user asks to write/mutate production data, or when a task needs a real prod DB connection. Also use when adding a DB model/table/migration, when applying a migration to a hosted project through a dashboard or MCP tool (Supabase apply_migration, Neon, PlanetScale) — see "Migrations applied by a hosted tool" — or when setting up a CI guard that enforces a per-table invariant (RLS enabled, tenant column, required index) — see "Enforce schema invariants in CI". Assumes a Neon/Vercel-style setup with the platform CLI, but the protocol generalizes.

Router for sourcing UI components — picks ONE source among the shadcn/ui registry (MCP or CLI), 21st.dev magic-mcp, and Ant Design, so a component need lands on one system instead of a hand-rolled lookalike or a mixed component zoo. Read this FIRST whenever a task needs a UI component or section — a button, dialog, data table, form, pricing section, hero, navbar, dashboard widget — or when the user asks to "add a component", "build a section", "use shadcn", "find a component for", "תוסיף קומפוננטה", "תבנה סקשן", "צריך טבלה/טופס/דיאלוג", "חפש קומפוננטה". Not for designing a brand or page from scratch (that is design work, not component sourcing), not for choosing the project's framework, and not for installing/configuring the tools themselves.