Index of every page on this site: llms.txt
All skills
Run it yourself

Build a complete Bolt CMS 6 website from a supplied design — Figma (Figma MCP), Pencil / pen.dev (.pen via Pencil MCP), exported images or PDF, an existing website, or a brief — by running the whole WORKFLOWS.md (Phases 0–8, all 20 skills) autonomously. Reads the design first, asks every question the workflow needs in one upfront intake, writes the site plan, then builds, verifies, commits gate by gate and keeps a resumable progress log. Use when asked to create, build or implement a website (or "make this design into a Bolt site").

  • 6 files
  • 46 KB
  • Updated 2 weeks ago
  • GitHub
Use this Skill: https://skilld.dev/gh/vondry/bolt-skills/bolt-build

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈139 tokens for metadata: the name and description. ≈3.3k when used: this file. ≈5.9k more on demand in 3 files.

Description uses 6.9% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈551 characters, or 6.9%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Build a Bolt CMS Website from a Design (bolt-build)

WORKFLOWS.md describes what to build in which order; this skill is how an agent runs it end to end with as little interruption as possible:

  1. read the design before asking anything, so most answers are proposals to confirm;
  2. ask all remaining questions in one upfront intake;
  3. get one approval (Checkpoint 0), then run Phases 1–8 alone — choosing documented defaults, logging every decision, fixing failing gates, committing after each gate;
  4. keep a progress log in the project's AGENTS.md, so any later session resumes where the last one stopped;
  5. finish with a full verification, an independent audit and a handover list.
File Use
references/intake.md Every question the 20 skills need, in five rounds, with defaults and the phase that uses each answer
references/design-sources.md What to extract from Figma MCP, Pencil MCP, images/PDF, an existing site or a brief; token mapping; visual QA loop
references/site-plan-template.md Site plan, build progress, decisions log, open issues and handover — written into the project's AGENTS.md
scripts/preflight.sh Tools, PHP extensions, Bolt version, console, database, git, local URL — before Phase 1
scripts/config-notices.php The admin dashboard's Bolt Configuration Notices from the command line — must be clean before any content is entered (Phase 5 gate, Phase 7)

The phase steps themselves are not repeated here: for every phase follow WORKFLOWS.md (Integration Contract, the phase section, its gate) and the SKILL.md of each skill it names. $SKILLS below is vendor/tomvondracek/bolt-skills/skills (or a checkout of this repository).


Step 0 — New build or resume?

Read the project's AGENTS.md. If it has a ## Build progress table, this is a resume:

  • re-read ## Site plan, ## Decisions log, ## Open issues and var/bolt-build/design/design-notes.md;
  • git log --oneline -15 and git status to see where the last session stopped;
  • continue at the first phase that is not done — do not repeat the intake, do not re-ask answered questions.

Otherwise start at Step 1.

Step 1 — Preflight

bash $SKILLS/bolt-build/scripts/preflight.sh --project=<dir> [--base-url=<local URL>]

Resolve blockers with the user now (missing Node, PHP extension, wrong Bolt version) — they would otherwise stop the run in Phase 4. Warnings are handled by Phase 1 or recorded in the site plan.

Step 2 — Read the design

Follow references/design-sources.md for the source the user gave. Produce, before any question:

  • the inventory (pages, sections, breakpoints), the content model proposal (ContentTypes, fields, menu profile, features), the token table, the component list, extracted copy and assets;
  • reference screenshots and design-notes.md in var/bolt-build/design/.

If the design tool does not respond (MCP server not connected), say so and ask the user to start it or to export images — never reconstruct a design from guesses.

Step 3 — Upfront intake

Ask the rounds in references/intake.md:

  • Round 1 (environment) and Round 2 (structure — mostly the design proposal to confirm) first;
  • Round 3 only for gaps the design left; Round 4 operator facts; Round 5 autonomy.

Use the agent's structured-question tool, where it has one, for multiple-choice items (recommended default first), and one message for free-text facts. Do not drip-feed questions across phases. What the user leaves unanswered gets the default and a Decisions log entry; operator facts (legal identity, IDs, retention) stay [DOPLNIT] and go to the handover list — never invented.

Secrets: never ask the user to paste passwords into the chat. Account creation is interactive, so the user runs it themselves in their own terminal (in Claude Code: type the command with the ! prefix, which runs it in the user's session): on a new project php bin/console bolt:setup --no-fixtures (migrations, APP_SECRET, first account — run it once: it resets APP_SECRET and writes a new migration every time); on an existing database php bin/console bolt:add-user --admin. bolt:setup writes APP_SECRET into the committed .env, which is where it belongs (bolt-security-hardening Step 1); .env.local only switches the machine to dev.

Step 4 — Site plan and Checkpoint 0

Write ## Site plan, ## Build progress, ## Decisions log, ## Open issues and ## Handover into the project's AGENTS.md (references/site-plan-template.md). Show the plan — with the design direction, the page/URL map, the ContentTypes and the menu — and get one approval. This is the only mandatory stop. Further checkpoints only if the user chose them (A1).

Step 5 — Run Phases 1–8 autonomously

For each phase: set its row to in progress, follow the WORKFLOWS.md section and each named skill's SKILL.md, run the phase's verifications, fix, commit, set the row to done, move on. Where the design and the intake answers flow in:

Phase Input from the intake / design Concretely
1 E1, E3, E6, E7, E9, E11, E13, O10, S1 composer create-project bolt/project <dir> (new); SQLite from Bolt's default .env for the local build (no database configuration; the production database of E13 goes into .env, which is uploaded); the user runs php bin/console bolt:setup --no-fixtures in their terminal (it prompts for the first account; never Bolt's sample fixtures); admin user (Step 3 secrets rule), git init + baseline commit (with /var/bolt-build/ in .gitignore), composer require --dev tomvondracek/bolt-skills; timezone, app_locales, API disable or lock-down, code quality for the chosen CI, bolt-security-hardening configuration (.env defaults, trusted hosts from E7)
2 E5, E6, E12 Apache vs. Nginx branch of bolt-htaccess / bolt-asset-caching; nelmio headers and baseline CSP; admin restriction if E12 asks for it
3 S6 keep / remove ImageDimensionsExtension (gallery), FileSizeExtension (downloads); note removals as expected warnings
4 E8, D1–D5, O1, O9, L2, token table bolt-theme-init scaffold <theme>; theme.yaml contact, socials, logo, analytics:; bolt-menu --profile=<S4> --templates --force; tokens and self-hosted fonts from design-notes.md into css/tailwind/index.css; JS imports per features
5 S3–S5, S1, content model proposal contenttypes.yaml, static routes, menu.yaml with labels per locale, catalogs
6 frames, components, O2–O8, L1–L5, S7, S11 one orchestrator per page from its frame, containers for queries, page-scoped components; visual QA loop (design-sources.md) per page; --products=auto; locale switcher; helper pages with the metaphor; bolt-cookie-consent scaffold with --model --trackers --embeds --purposes from O2–O7; _privacy-form-note beside every form; embeds through _consent-embed; bolt-legal-cz scaffold with --operator --consumers --online-contracts --reviews from L1–L5, identity checked with verify.sh --ares, _reviews-notice under testimonials; optionally bolt-performance fragment caches for pages with a high query count; last, bolt-edit-links across the finished templates
7 S2, S8–S10, O11, O12, E11, copy first php $SKILLS/bolt-build/scripts/config-notices.php --project=. must exit 0 (resolve every notice, clear the cache, re-run); then records from the design copy or supplied content (through the admin in a browser if S10 allows), media-library uploads, the admin icon render check (bolt-contenttypes/scripts/admin-icons-browser.js, rule 11), translations, SEO per locale, llms.txt, npm run prod, assets_version, deployment checklist (7.5) for the hosting of E11
8 E4, E7 linters, verify-all.sh with BOLT_BASE_URL, smoke checks, manual checks with a browser tool, Lighthouse, then the independent audit (Step 6)

Autonomy rules

  • Stop only for blockers: missing credentials or access, an environment failure the agent cannot fix (database server down, network), a decision outside the site plan that changes architecture (a new language, hosting change, a feature not in the design), or anything destructive outside the project. Everything else: apply the documented default, log it, continue.
  • Decisions log: every default applied without confirmation and every judgement call (a field added, a breakpoint chosen, a token adjusted for contrast) — one dated line each.
  • Failing gate: diagnose with the skill's verify output and WORKFLOWS.md Troubleshooting, fix, re-run. After three failed attempts (A3): record it under Open issues with what was tried, mark the phase done with open issues, and continue only if later phases do not depend on the failing part; otherwise mark it blocked and stop with a report.
  • Verifications scheduled later (WORKFLOWS Appendix A) are not failures — do not chase them early.
  • Commits after every gate, in the project's commit convention (default chore(build): phase N — <summary>); never push unless E10 says so; never rewrite history.
  • Stay inside the project: no global installs, no changes to other directories, no deployment.
  • No fixtures, ever: the database is SQLite from Bolt's default .env, initialised once with bolt:setup --no-fixtures. Never run doctrine:fixtures:load or bolt:setup without --no-fixtures; never run make db-reset on a database with content (it drops everything). Schema changes after Phase 5 go through make db-update — on the local database only: it drops tables Doctrine does not map and marks migrations as run without running them; servers run doctrine:migrations:migrate (bolt-code-quality Step 4).
  • Operator facts are never invented (bolt-cookie-consent AG-01, bolt-legal-cz AG-02): [DOPLNIT] and the handover list. The operator type (L1) is always asked, never assumed.
  • Context discipline on long runs: re-read the phase section and the skill's SKILL.md at the start of each phase rather than relying on memory; keep design details in design-notes.md, not in the conversation. A new session resumes from the progress table.
  • Parallel agents only if the user allowed them (A4) — e.g. independent page templates in 6.1. Each gets the site plan, the Integration Contract and its frames; the main agent integrates, builds and verifies.

Step 6 — Final verification and audit

  1. Phase 8 of WORKFLOWS.md in full (8.1–8.5).
  2. An independent check of the finished site with the audit runner:
    bash $SKILLS/bolt-audit/scripts/audit.sh --base-url=<local URL>
    Every skill should be PASS, or WARN with a reason listed under Expected verification warnings. Fix what is not.
  3. Visual sign-off: final screenshots of every page at every design breakpoint and 360 px next to the references, saved in var/bolt-build/qa/.

Step 7 — Handover

Report to the user, briefly:

  • the local URL, admin URL, and how to start the site;
  • what was built (pages, ContentTypes, languages, features) and the final verification result;
  • Open issues and remaining visual differences (with screenshot paths);
  • the Handover list: every [DOPLNIT] and [OVĚŘIT] (privacy data, legal identity and consumer pages, tracker IDs, measured cookies), unconfirmed defaults, content the operator must supply;
  • what is outside the workflow: deployment, DNS, mail delivery for forms, production database.

Pitfalls

  • Asking question by question during the build. Everything answerable is in the intake; new questions mid-run are a sign a default should have been applied and logged.
  • Pasting generated code from a design tool. Figma's get_design_context returns React + Tailwind with arbitrary values; translate it into Twig with contract tokens and components, or the undefined-class check fails and the theme drifts from the tokens.
  • Design colours outside the contract. Map to roles (design-sources.md); a new colour name renders unstyled.
  • Google Fonts from Google's CDN violates the consent rules — self-host (bolt-tailwind-theme).
  • Iframes straight from the design (YouTube, maps) — always _consent-embed.twig.
  • bolt-menu scaffold without --force on a fresh bolt/project stops before copying templates (sample menu.yaml exists).
  • Seeding content before Phase 5 is final — every ContentType change after content entry means re-entering records.
  • Entering content while Bolt shows configuration notices. A reserved field name (icon, status, label / value inside a set …) or an inferred slug is invisible in the editor but breaks the templates, and renaming the field later drops its stored values. Run scripts/config-notices.php (or look at the dashboard) and resolve every notice before the first record — bolt-contenttypes rules 8 and 9.
  • Skipping the progress table update — the next session cannot resume reliably.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 083eb14. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 3 days ago.

Activeupdated 2 weeks ago

README badge

README badge for vondry/bolt-skills/bolt-build