GenerateBlocks Skills
LLM-optimized agent skills for GenerateBlocks V2/Pro that let AI coding assistants produce valid block markup you can publish to WordPress live over MCP.
- Source repo
- wpgaurav/generateblocks-skills
- Stars
- ★ 143
- Last updated
- 25d ago
- Primary language
- PHP
- FA score
- 66/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A WordPress developer building with GenerateBlocks V2/Pro who wants Claude Code to generate hero sections, pricing tables or FAQ markup directly
- Not a fit if
- WordPress sites that don't use GenerateBlocks or GeneratePress
- Projects needing interactions built on custom JavaScript rather than CSS
- Users without an agent-capable editor or CLI who only build pages by hand
- Website
- gauravtiwari.org ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-10-28
- Source review
- 66/100 · Some gaps
What does this agent do, and when should you use it?
GenerateBlocks Skills is an open-source collection of skill documentation for AI coding assistants, verified against GenerateBlocks 2.5.0 and Pro 2.8.0. It ships four skills: generateblocks-layouts (static sections, query loops, dynamic tags, ACF fields, Pro forms and interactive blocks), html-to-generateblocks, elementor-to-generateblocks and figma-to-generateblocks. Each skill is a slim SKILL.md entry point backed by routed reference files, validated examples, and a preflight.py validation script. The execution model is purely documentation-driven: the skills run no service themselves — assistants like Claude Code, Codex, Cursor, Windsurf and Gemini CLI read the rules, then output block-markup strings that are either pasted into the block editor's Code editor or published via an MCP plugin such as Site Agent that reads and writes raw post_content. The repo was verified in a live WordPress 7.1.3 editor, and its bundled build_css() compiler matches the native 2.5 Styles Builder byte for byte across 1,453 style objects. Plugin source is GPL-2.0-or-later; skill documentation is MIT.
The repo provides four skill directories (generateblocks-layouts, html-to-generateblocks, elementor-to-generateblocks, figma-to-generateblocks), each with a SKILL.md entry, references/ files (authoring-contract.md, recovery-rules.md, dynamic-tags.md, query-block.md, pro-forms.md, mcp-publishing.md and more) and examples/. After reading a skill, an AI assistant produces GenerateBlocks V2 markup: the four core blocks generateblocks/element, /text, /media, /shape plus the generateblocks/query family, with styling in editable styles attributes compiled into local css. Publishing has two paths: manual paste into the block editor's Code editor (Ctrl+Shift+Alt+M), or automation through an MCP server that treats post_content as an opaque string (the author's Site Agent plugin is the verified route) combined with verify_roundtrip.py checks. install.sh deploys skills interactively to per-tool directories; importable/ offers .skill/.zip archives. Run preflight.py <file> --id-scope <SCOPE> before each write.
- A WordPress developer building with GenerateBlocks V2/Pro who wants Claude Code to generate hero sections, pricing tables or FAQ markup directly
- A team migrating Elementor layouts or Figma designs to clean GenerateBlocks blocks
- A site maintainer turning ACF repeater fields (e.g. team_members) into dynamic grids or query loops
- A publishing workflow where an agent writes generated blocks into a staging or live site via the Site Agent MCP plugin
- A builder needing full-site templates such as GeneratePress Elements loop templates and FSE block themes
- A frontend developer curating AI skills for the GenerateBlocks ecosystem alongside Bricks and Kadence
How do you install or deploy this agent?
Quick install via the bundled multi-tool installer:
git clone https://github.com/wpgaurav/generateblocks-skills.git
cd generateblocks-skills
chmod +x install.sh
./install.shChoose tools interactively, or install in one shot:
./install.sh --all
./install.sh claude
./install.sh cursor codexClaude Code skills go to ~/.claude/skills/, Cursor to ~/.cursor/skills/; for Codex's native skill-directory installation see the wiki Installation guide. Alternatively, skip the installer and reference skill files directly in Claude Code/Cursor/Windsurf, or upload a .skill/.zip archive from importable/.
How do you use this agent?
Once installed, issue prompts that reference the skill, e.g.:
Read skills/generateblocks-layouts/SKILL.md and create a testimonial slider with 3 cardsor a typical publishing flow:
Read skills/generateblocks-layouts/SKILL.md, then generate a testimonial grid with 3 cards and insert it as a new section at the top of /about on my staging site.Paste generated markup into the block editor's Code editor (Ctrl+Shift+Alt+M) as a user with unfiltered_html, and never click "Attempt Block Recovery". For automation, read references/mcp-publishing.md first: any MCP server must read and write raw post_content as an opaque string, and verify read-backs with verify_roundtrip.py. Always validate on staging first.
What are this agent's strengths and limitations?
- Verified against GenerateBlocks 2.5.0/Pro 2.8.0 in a live editor; build_css() matches the native Styles Builder byte for byte (1,453 style objects)
- Fine-grained references: exact syntax for all 27 dynamic tags, complete cause catalog for 'Attempt Recovery' errors, JSON attribute-order rules
- One-command install.sh covers Claude Code, Cursor, Windsurf, Codex, Gemini CLI, Copilot, Cline/Roo and Aider
- Ships preflight.py and verify_roundtrip.py scripts that catch markup issues pre-write and name specific corruption after MCP round trips
- Locked to the GenerateBlocks/GeneratePress ecosystem; useless for sites on other builders
- Interactions are CSS-only plus GB Pro's built-in JS blocks; no custom JavaScript, no external URLs, placeholder images only
- UI-managed settings — condition rules, form actions, overlay triggers, display rules — must be configured manually in wp-admin
- MCP publishing carries real risk: a write returning 200 can still corrupt blocks, requiring byte-level read-back verification
How does this agent compare with similar options?
Siblings from the same author: bricks-skills (Bricks Builder JSON), kadence-skills (Kadence Blocks) and WordPress-skills (125+ WordPress/design/SEO skills). For MCP publishing, Site Agent is the recommended route for GenerateBlocks, with WordPress MCP Adapter, Novamira, WPVibe and Respira as documented alternatives.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| GenerateBlocks Skills This agent | 66 · Some gaps | Agent plugin / skillFree | ★ 143 | 25d ago | PHP | Codex · Claude Code |
| TermFold Debian Coding Terminal | 62 · Some gaps | Desktop appFree + model costs | ★ 36 | 1mo ago | Kotlin | Codex · Claude Code |
| Agent Graph | 53 · Major gaps | Self-hosted serviceFree + model costs | ★ 239 | 5d ago | Python | — |
| Mirafold | 87 · Good | CLIFreemium | ★ 20 | 2mo ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows read-only defaults, file-based output, and live writes gated behind mcp-publishing.md, content.raw snapshots, raw readback, and an explicit 'conversion does not authorize publication' rule; shared Global Styles/Design Tokens require explicit opt-in, and Site Agent's PHP/file tools ship disabled. Deductions: sensitive-data handling only mentions Application Passwords and unfiltered_html with no systematic credential guidance; install.sh is not in the packet so its behavior cannot be statically audited; attribution is clear though publisher identity is unverified (no deduction for that).
SKILL.md files, AGENTS.md and CLAUDE.md agree on versions, loading rules, and validation flow; standalone converter bundles embed the layout dependency to avoid broken paths; recovery-rules.md and verify_roundtrip.py name specific corruption types. Deduction: consistency rests on declarations — the scripts themselves are not in the packet, so static review can only credit the descriptions.
Covers static sections, query loops, ACF, conditions, Pro forms, full-site templates, and three converters, with a multi-client install matrix that honestly caveats 'not a guarantee of automatic loading'. Capability boundaries (no external URLs, no custom JS, placeholder images, version-dependent Pro features) are explicit and actionable — full marks. Deduction: trigger lists exist but no conflict-resolution rule between overlapping triggers is shown.
Strong information architecture: _index.md router, compact authoring-contract.md, on-demand reference loading (3). Limitations are concrete. Deductions: no repository LICENSE (only Site Agent is labeled GPL-2.0-or-later); maintenance is a single author with no governance path; install.sh, build-bundles.sh and any CHANGELOG are outside the packet, so install and change history depend on README/Wiki links.
The delivery contract is directly usable: write files not chat, run preflight.py, confirm in the real editor, verify raw readback; on-demand references control token cost. Deduction: real-world value depends on editor validation on the target site, which static evidence cannot establish; the packet spends notable space cross-promoting the author's other products (Site Agent, Stackless, GT Performance).
Verification claims are specific enough to check (1,453 style objects, 800 untuned-seed samples, native_css_2_5_0.json fixture, dated environments), with companion scripts and examples. Deductions: all verification is self-reported with no third-party or CI corroboration (cross-source 1); the 2026-dated live verification cannot be confirmed in static review; README largely separates facts from inference (the Figma skill even mandates distinguishing measured vs inferred), but promotional content is interleaved with factual claims.
- No repository license is stated; confirm licensing terms before enterprise use.
- install.sh is not in the review packet: read it before running; note it overwrites target rule files (README advises preserving unrelated instructions).
- Live WordPress writes are high-risk: validate on staging first, always snapshot and verify raw readback, and never let the agent write to production without authorization.
- All 'verified' claims are self-reported (including 2026-dated runs); re-verify against your actual target versions before relying on them.
- README promotes the author's own products; when choosing an MCP route, compare the neutral alternatives listed (official mcp-adapter or plain REST).
FAQ
Is there a cost?
How does generated markup reach a real site?
Which AI tools are supported?
Why does a block show 'unexpected or invalid content'?
Can I let the agent write to production?
Is GenerateBlocks Skills open source?
Is GenerateBlocks Skills safe to use?
Is GenerateBlocks Skills still maintained?
Who makes GenerateBlocks Skills?
What is GenerateBlocks Skills's official website?
Thanks — we use this to improve the page.