NextBoard: an AI agent that turns a product idea into a reviewable PCB schematic plan

The gap NextBoard targets: requirements that never become a reviewable plan

Hardware planning usually starts as a sentence. Someone wants a drone, a FOC controller, or an e-paper vocabulary card, and the first real work is turning that sentence into a frozen requirement set, a shortlist of architectures, a part list with prices, and a risk register. NextBoard is aimed at that stage. The README describes it as an AI Agent for PCB solution design that accepts a product requirement and outputs a plan that can be reviewed and implemented.

The intended user is an engineer or a small team that already has an AI coding session open and wants the early design exploration structured rather than ad hoc. The README's own examples are concrete: design a drone, design a FOC controller, design an e-paper card for memorizing words. Nothing in the repository suggests it replaces a hardware engineer. It produces a plan, and the plan is meant to be reviewed.

Seven stages, three architecture candidates, five gates

The workflow is the substance of the project. According to the README, the skill runs a structured seven-stage process: requirement freeze, architecture candidates, system decomposition, component selection, output generation, review, and verification gating. At the architecture stage it compares three candidate directions: domestic-first, overseas mainstream, and a hybrid compromise. That three-way split is the most opinionated design decision in the repository, and it is clearly aimed at sourcing rather than at electrical performance.

Five verification gates interrupt the flow to catch quality problems stage by stage. A separate review agent scores the result across five dimensions: completeness, risk, implementability, cost, and verification coverage. The references directory backs this up with design-workflow.md, verification-gates.md, review-checklists.md, sourcing-and-risk.md and domestic-sources.md. The gates are not decorative: the validation script checks that gate files are non-empty and that the output template covers Gate 4.

Outputs listed in the README are markdown, datasheets, a footprint library, and a PDF plan. The README does not describe how the footprint library is generated or what format it uses, which is the thinnest part of the documented output set.

Installing NextBoard and running the hardware-solution skill

NextBoard installs as a plugin or skill for Claude Code, Codex and Cursor. The README recommends the interactive script first, which reports current install status and then asks which of the three modes you want. Run it from a clone of the repository.

git clone <NextBoard-repo-url>
cd NextBoard
./scripts/install.sh

If you already know the mode, the script takes flags directly. Global install copies the skill and the review agent into your home directory so every project can call it; the README notes that hooks do not take effect in this mode because the plugin context is missing.

./scripts/install.sh --global --platform claude
./scripts/install.sh --global --platform codex

For development or a one-off session, load the plugin directory instead. This is the only mode where skill, agent and hooks are all active, and it requires passing the path on every Claude Code launch.

claude --plugin-dir /path/to/NextBoard

Marketplace install is the distribution path. The README gives two commands, and notes it requires GitHub access.

claude plugin marketplace add LeoKemp223/NextBoard
claude plugin install nextboard-hardware-solution

After any install, restart the AI IDE session so the client reloads the skill list, then invoke the skill by typing its name in the session. The README's quick start shows a single command followed by your product requirement.

$hardware-solution

What you should see is the agent entering the design flow: confirming requirements, proposing architecture candidates, and moving toward a BOM and a plan. If the skill does not appear, the README's status command tells you whether the install registered.

./scripts/install.sh --status

Where NextBoard stops: no layout, no netlist, and a hook asymmetry

The most important limitation is scope. NextBoard outputs a module schematic plan, a BOM, datasheets, footprints and a PDF. It does not perform PCB layout, and the README never claims it does. If your next step is routing a board, this tool finishes before that step begins.

The second limitation is install-mode asymmetry, and it is documented rather than hidden. Hooks only fire in project-level plugin mode, because .claude-plugin/plugin.json declares "hooks": "./hooks/". Global installs get the skill and the review agent but no hooks. Codex does not support plugin mode at all, so Codex users are limited to global install. Cursor's global path is a manual cp of the skill. Anyone who assumes all three platforms behave identically will be wrong.

The third is data provenance. The README's output quality rules state that component parameters must come from datasheets or distributor pages and must not be recalled from memory. That is a rule imposed on the agent, not a guarantee enforced by the code. The validation script checks structure, content consistency and anti-patterns such as vague wording and leftover placeholders. It does not verify that a resistor value matches a datasheet. Treat the BOM as a draft to be checked, not as a verified artifact.

Finally, the repository has no releases and the README does not document rollback. Uninstall is covered, but reverting to a previous working version means checking out an earlier commit yourself.

Compared with a plain LLM prompt or an EDA tool's part search

The obvious alternative is asking a general model directly: paste your requirement into a chat and ask for a BOM. The difference is process. A single prompt has no requirement freeze, no three-way architecture comparison, no gate that blocks output until the risk list is non-empty, and no separate reviewer scoring completeness and cost. NextBoard encodes those steps in SKILL.md and the reference documents, so the structure survives across sessions and across users.

A second alternative is the part search built into distributor sites or EDA tools. Those are strong at what they do: parametric search over real inventory with live pricing. NextBoard's README claims it calculates prices and assesses supply chain risk, but the sourcing-and-risk.md and domestic-sources.md references are guidance documents, not an inventory API. If your question is which 10k resistor is in stock today, a distributor search answers it with authoritative data. NextBoard's value is upstream of that, deciding which parts the design needs and why.

The honest framing is that NextBoard is a planning layer. It competes with a blank page and a scattered set of browser tabs, not with KiCad.

Maintenance, upgrade cost and the MIT licence

The repository is not archived, and the last push was on 2026-05-05. That is roughly four months before today, so the project is not in the archived category, but there are no releases to pin to and the README does not describe a versioning scheme beyond a .version-bump.json file at the repository root.

Upgrade cost is low and documented. For a global install, pull and re-run the installer; for plugin-dir mode, pull and the next launch picks up the new files.

cd NextBoard && git pull
./scripts/install.sh --global --platform claude

The real upgrade risk is content drift. The skill is a set of markdown instructions, so a change to SKILL.md or a reference document changes agent behaviour without any version number moving. The README's contribution rules acknowledge this: changes to the skill or reference docs require a before-and-after comparison, and new reference documents must be linked from the SKILL.md flow section. Run the validator after any pull.

python3 tests/validate.py
python3 tests/validate.py --installed

The licence is MIT, stated in the README. That permits commercial use and modification with attribution and no warranty. It also means the project's dependency on Claude Code, Codex and Cursor terms sits outside the licence, and nothing in the repository addresses what happens to generated designs under those platforms' terms.

Editorial conclusion

Adopt NextBoard if you already work inside Claude Code, Codex or Cursor and want a structured first pass at architecture candidates, part selection and supply chain risk before anyone opens a schematic editor. Skip it if you need netlist-accurate schematics, PCB layout, or a tool that runs without an AI IDE session. Before relying on it, run python3 tests/validate.py against your installed copy and check whether hooks are actually active in your install mode, because global installs do not include them.

Frequently asked questions

Official sources

1. Issues
2. LeoKemp223/NextBoard on GitHub
3. README