dapp-builder

Freenet Decentralized Application Builder

Build decentralized applications on Freenet following the architecture patterns established in River (decentralized chat).

How Freenet Applications Work

Freenet is a platform for building decentralized applications that run without centralized servers. Apps store and exchange data through a global, peer-to-peer Key-Value Store shared by every Freenet node.

The keys in that store are not arbitrary strings — they're derived from small pieces of WebAssembly called contracts that define how each value is allowed to change. The next two sections introduce the kinds of components that make up a Freenet app, then explain exactly how contract keys are formed and why that makes the system trustless.

The Three Kinds of Components in a Freenet App

A Freenet app is built from three kinds of components — contracts, delegates, and a UI. Most non-trivial apps have multiple contracts and multiple delegates, each handling a different concern.

1. Contracts (Network State)

A Freenet app typically has one or more contracts, each defining a different kind of shared state. River has a single room contract today, but a more complex app might have several (e.g. rooms, user profiles, invitations, search indexes), and each one is a separate contract crate that compiles to its own WASM.

• Role: Closer to a database table than a database. The contract WASM defines the schema (state shape) and the rules for validation and merging. Each instance of the contract — there can be many — behaves like an independent row in that table, so a chat app can have thousands of "room" rows all governed by the same room-contract WASM. (How rows are addressed is covered in How Contract Keys Work below.)
• Location: Runs on the public network (untrusted peers).
• Functionality:
  • Defines what state is valid
  • Defines how state can be modified (validate / update / summarize / delta)
• State: Holds the actual application data for that instance (arbitrary bytes).
• Constraint: Cannot hold private keys or secrets — all state is public unless encrypted by the client.

2. Delegates (Local Trust Zone)

A Freenet app may have one or more delegates, each handling a different local responsibility — key management, secret storage, background sync, notifications, and so on. Delegates are the local counterpart to contracts: where contracts hold shared state on the network, delegates hold private state on the user's device.

• Role: Trusted middleware between the user and the network.
• Location: Runs locally on the user's device, inside the Freenet kernel.
• Functionality:
  • Trust Zone: Safely stores secrets, private keys, and user data
  • Computation: Performs signing, encryption, and complex logic before publishing to the network
  • Background Tasks: Can run continuously to monitor contracts or handle notifications even when the UI is closed

3. The User Interface (Frontend)

A single UI typically talks to all of an app's contracts and delegates.

• Role: Interaction layer for the user
• Location: Web Browser (SPA) or native app
• Functionality:
  • Connects to the local Freenet Kernel via WebSocket/HTTP
  • Built using standard web frameworks (Dioxus, React, Vue, etc.)
  • Agnostic to underlying P2P network complexity

"Native app" above means desktop. Freenet does not currently support running a full node on mobile devices. Do not recommend or generate a production mobile wrapper without clearly warning about likely bandwidth, battery, thermal, CPU, and background-execution problems. Treat any such work as experimental, require explicit resource measurements before calling it viable, and do not represent it as an official Freenet client without approval from the Freenet Project.

How Contract Keys Work (and Why Freenet is Trustless)

Now that contracts have been introduced, here's how they're addressed in the network.

The key for a piece of data is derived from the cryptographic hash of the contract's WebAssembly (WASM) code, combined with a set of contract parameters that identify a specific instance.

• The WASM hash ties the identity of the data to its logic — change the code, and the key changes.
• The parameters distinguish independent instances of the same contract code. Tying back to the database-table analogy: the WASM is the table schema, and each parametrized instance is a row with its own key and its own state.
• This is what makes the network "trustless" — you don't have to trust the peer that holds the data, because the data is self-verifying against the contract code referenced in the key.

Data Synchronization & Consistency

Freenet solves "Eventual Consistency" using a specific mathematical requirement:

Commutative Monoid: The function that merges updates must be a commutative monoid.

• Order Independent: It shouldn't matter what order updates arrive in
• If Peer A merges Update X then Y, and Peer B merges Update Y then X, they must end up with the same result

Efficiency: Peers exchange Summaries (compact representations) and Deltas (patches/diffs) rather than re-downloading full state.

Requirement: get_state_delta must not ship state to a peer that already has it. When the requester's summary shows it holds everything you have, the delta carries no information, so it must not contain the state or approach the state's size. It should be a literally empty StateDelta (vec![]), which is the unambiguous "converged" answer and what freenet-scaffold produces for you; a few tens of bytes of encoding framing from serializing an all-empty struct is acceptable. What matters is delta size relative to state size: 20 bytes against a 500 KB state is fine, a state-sized delta is a broken delta mechanism that re-ships everything on every reconciliation, forever. Your summary must likewise be far smaller than your state. Core is adding a probe for contracts that get this wrong, and it currently costs the network real bandwidth. Full detail, code shapes, and a test are in references/contract-patterns.md → "The Delta to an Up-to-Date Peer".

Summaries must serialize deterministically

Use deterministic maps in your Summary type: BTreeMap, never HashMap. A HashMap serializes in nondeterministic order (ciborium), so two identical states can summarize to different bytes and core's byte-level convergence check misfires — spurious heals, or missed ones. The same caution applies to any map inside whatever summarize returns.

Beyond that, make sure your state genuinely converges through summarize / delta / apply, and test that it does, rather than assuming a live broadcast reaches every peer.

Previously documented here as a live limitation, now fixed. freenet/freenet-core#4857 ("State updates permanently lost for rarely-changing fields") is CLOSED. A ContractQueueFull drop was silent, and the sender cached its own summary as the receiver's on send-Ok, so it believed the peer was current and never re-sent — leaving rarely-changing fields (config, permissions, ban lists) diverged until the ~5-minute InterestSync heartbeat happened to correct them.

The shipped fix has the queue-full receiver emit a ResyncRequest, which makes the sender clear its poisoned summary and re-send full state. It is throttled to one per (contract, peer) per 30s, because #4251 showed that one request per dropped delta amplifies into a full-state storm onto the same saturated queue; #4862 hardened it against bridge backpressure. See RESYNC_REQUEST_MIN_INTERVAL in crates/core/src/ring/interest.rs.

Do not design around multi-minute staleness on rarely-changing fields, and do not treat a ban list or permission field as needing to ride alongside a frequently-changing one. The earlier guidance to do so is retired.

Advanced Capabilities

• Subscriptions: Clients can subscribe to contracts and get notified of changes immediately (real-time apps)
• Contract Interoperability: Contracts reading other contracts' state is planned but not yet implemented

Development Workflow

Follow these phases in order.

Already shipped v1 and here to UPGRADE? (bump freenet-stdlib, ship a new contract/delegate version, or fix a bug that re-keys the WASM) — go straight to references/upgrade-and-migration.md → "Upgrading a Freenet dApp — the painless path", the single start-to-finish playbook. A routine WASM/stdlib bump is low-risk and mechanical when you designed for it at v1, not "recreate everything and all invites die" — River's live 0.6→0.8 stdlib re-key (verified 2026-07-12) auto-migrated every room on refresh, kept every invite and the 78-member Official room intact, and needed no recreation. The phases below build a new dApp; the playbook ties the upgrade steps together (v1 design precondition → reproducible builds → register the outgoing hash → freenet-migrate → publish → do NOT recreate instances or warn of dead invites).

Phase 1: Contract Design (Shared State)

Start by listing each kind of shared state your app needs — each kind becomes its own contract crate. Then design each one in turn using the questions below.

Key questions (per contract):

• What data must all users see consistently for this concern?
• How should conflicts be resolved when two users update simultaneously?
• What cryptographic verification is needed?
• What are the state components and their relationships?
• What parameters distinguish one instance from another (e.g. room owner key, profile owner key)?
• If users reference each other (messaging, contacts, profiles), what is the user-facing identifier? It should be short, self-certifying, and stable across WASM upgrades — derived from a key, never a contract key. See identity-and-addressing.md.
• Can strangers write to this contract? If so, what stops one attacker minting ten thousand keys and flooding it? There is no server to host a CAPTCHA, so the mechanisms available are proof-of-work and ghost keys (a blind-signed certificate proving an anonymous donation — a cost the attacker cannot beat with better hardware, and not burned as waste heat). The recommended shape is both: proof-of-work as the always-sufficient default, with a ghost key offered as a way to skip the wait, surfaced while the grind is running and the user is blocked anyway. That lets you set difficulty by what deters an attacker rather than by what your slowest device tolerates, while never pricing anyone out. See identity-and-addressing.md → "Cryptographic CAPTCHA". Present the choice to the developer rather than picking silently: ghost keys cost their users money, that money funds Freenet, and the mint is centralized (verification is not). Those are product and architecture decisions, not technical details to settle on the developer's behalf.
• How large can this contract's state realistically grow? Keep each contract instance's state small — target well under 4 MB, not just under the host's 50 MiB hard cap. A GET transfers the entire state before the UI can render anything, so state size is felt directly as load latency, and validate_state/update_state/summarize_state WASM execution cost scales with it too. If a kind of data can grow without bound (message history, uploaded files, a membership list that only grows), don't let one contract instance absorb all of it — shard by the natural unit of write concurrency instead (one contract per room, per user, per time-window, per shard-key, etc.), so each instance stays small regardless of how large the dataset gets in aggregate. See state-authorization-patterns.md → "State Size Budget".

Implementation steps:

1. Define state structure using #[composable] macro from freenet-scaffold
2. Implement ComposableState trait for each component
3. Implement ContractInterface trait for the contract
4. Ensure all state updates satisfy the commutative monoid requirement, and that get_state_delta returns a negligible delta (ideally zero bytes, never state-sized) when the requester's summary already matches your state (see contract-patterns.md)
5. Every field in state must be covered by a cryptographic signature -- contracts run on untrusted peers who can modify unsigned fields. Write a test for each signed field verifying that tampering causes verification failure. See contract-patterns.md for versioned signature patterns when adding fields later.
6. Plan contract upgrade from v1 — it's low-risk and mechanical when you design for it. A WASM change moves the contract key, but if you anchor your app's durable references on a stable identity anchor that is independent of the WASM — never the raw contract key — the upgrade is transparent to users: state migrates itself on next load, and every reference derived from that anchor (invites, share links, membership, external services) survives the re-key because the client re-derives the new contract key from the unchanged anchor, not from a stored contract key. What the anchor is depends on the app (an owner/user key — e.g. River; a fixed namespace/singleton params; a DID; or an index contract mapping a stable name → current key); the options are in upgrade-and-migration.md step 1. Invites and links do not die on an upgrade — River's 0.6→0.8 re-key on the live network kept every room and invite. Recreation is only for deliberately changing the app's identity anchor (e.g. rotating a compromised owner key), never for a routine contract/stdlib bump. The shipped baseline (River #292, Delta) is a backward probe from a committed legacy-code-hash registry: reconstruct each predecessor key from (stable params ‖ old code_hash), GET the old state, fold it forward, and re-PUT under the current key — permissionless because the successor's validate_state re-verifies every byte. The one required operational step is registering the outgoing code hash in the registry before the WASM changes, then republishing. An author-signed OptionalUpgrade pointer is an optional straggler courtesy on top, not the mechanism that moves state. The reusable freenet-migrate crate (0.3.0 on crates.io, with freenet-migrate-build 0.2.0) packages this baseline and owns the probe decisions in a sans-IO driver; it is what River's contract-migration path runs in production (browser UI and riverctl), and existing apps adopt it without a rewrite because its build codegen reads their existing [[entry]] registries and emits view consts matching the hand-rolled shapes (freenet/river#434, #436, #437). See contract-patterns.md → "Contract WASM Upgrade & State Migration". For the cross-cutting operational discipline that keeps the migration itself from losing data (resumable, idempotent, non-destructive, regression-gated, observable), see upgrade-and-migration.md.
7. Read state-authorization-patterns.md before designing the second iteration. It captures cross-cutting patterns (per-item vs bundled signatures, replay protection via monotonic counter / tombstones / cross-context binding, signed-payload hygiene, time::now() gotchas, related-contracts limits, wire-format stability) that bite on every contract beyond the trivial.

References:

• references/contract-patterns.md — ContractInterface, commutative monoid, composable state, basic signatures.
• references/state-authorization-patterns.md — authentication, replay protection, signed-payload hygiene, time, related-contracts, wire-format stability, common pitfalls.
• references/identity-and-addressing.md — short self-certifying user-facing addresses, keeping large (post-quantum) keys out of identifiers, identity that survives WASM upgrades, and blocking bots without a server (ghost keys vs proof-of-work).

Phase 2: Delegate Design (Private State)

Determine what private data each user needs stored locally and split it across delegates by responsibility (e.g. one delegate per trust boundary or per long-running background task). Most apps need at least one delegate; many need several.

Key questions (per delegate):

• What user-specific data needs persistence? (keys, preferences, cached data)
• What signing/encryption operations are needed?
• What permissions are needed for sensitive operations?
• Does a platform delegate already do this? Delegates are callable across apps, so some responsibilities are worth borrowing rather than building. ghostkeys is the one that exists today: it holds the user's ghost-key identities and signs on your behalf, and it renders the user's allow/deny prompt itself, so you implement no permission flow. See identity-and-addressing.md → "Cryptographic CAPTCHA" for what it is for, how it pairs with proof-of-work as an escape hatch rather than replacing it, and the two caveats worth relaying to the developer: the mint is centralized, and Freenet has a funding interest in you choosing it.

Implementation steps:

1. Define request/response message types
2. Implement DelegateInterface trait
3. Handle secret storage operations (Store, Get, Delete, List)
4. Implement cryptographic operations (signing, encryption)
5. Design for secret migration from v1 -- when delegate WASM changes, the delegate key changes and all stored secrets become inaccessible. There is no ExportSecrets handler (an earlier misconception): River's real mechanism messages each old delegate key via DelegateRequest::ApplicationMessages, re-running the old WASM to read its secrets, and folds the signing keys forward (encryption secrets are re-derived). Keep a committed registry of old delegate keys and migrate promptly — the re-run breaks after a stdlib/ABI bump (freenet/river#204). See delegate-patterns.md for the mechanism; freenet-migrate codifies the delegate registry and build codegen, though the delegate secret carry-forward itself still runs app-side today (the crate's field-deployed path is the contract side, freenet-core#2776). See upgrade-and-migration.md for the operational discipline (resumable/interrupted-migration recovery, migration telemetry, and the upgrade test harness).

Reference: references/delegate-patterns.md

Phase 3: UI Design

Build the user interface connecting to contracts and delegates. Two approaches:

Option A: Dioxus (Rust → WASM)

Best for: teams already in Rust, complex state logic shared with contracts.

Implementation steps:

1. Set up Dioxus project with WASM target
2. Implement WebSocket connection to Freenet gateway
3. Create synchronizer for contract state subscriptions
4. Implement delegate communication for private storage
5. Build reactive UI components
6. Vendor your stylesheets, fonts, and scripts. The gateway serves every webapp under a same-origin CSP — CDN <link> / <script> tags from cdn.jsdelivr.net, cdnjs.cloudflare.com, fonts.googleapis.com, etc. are blocked in production even though they work in dx serve / vite dev. See references/ui-patterns.md "Gateway CSP: Vendor Your Assets".

Option B: TypeScript + Vite

Best for: web developers, faster iteration, familiar tooling (npm, SCSS, etc.).

Implementation steps:

1. Set up Vite project with @freenetorg/freenet-stdlib (TypeScript package)
2. Use FreenetWsApi class for WebSocket connection (handles FlatBuffers serialization)
3. Pass empty string auth token to FreenetWsApi constructor (sandbox blocks cookie reading)
4. Use Vite define to inject contract hashes and delegate key bytes at build time
5. For delegate communication, dynamically import internal FlatBuffers types (ClientRequestT, ApplicationMessagesT, etc.)
6. Build reactive UI with vanilla TS, or any framework (React, Vue, Svelte)

Validate the UI in a real browser (both options)

A Freenet UI's real render path only runs in a browser. A Dioxus UI ships as a WASM bundle, so rendering a component tree to a string in a Rust test does not exercise the compiled bundle, its event handlers, or its asset paths, and both options reach the node over a WebSocket that unit tests never touch. Drive the UI with Playwright (or equivalent browser automation) from the first screen onward, not only at release time. Treat "I built the component" as unfinished until a browser has loaded it and a script has clicked through it.

1. Serve the UI locally (dx serve for Dioxus, vite dev for TypeScript) and drive it with Playwright against mock or offline data, so render correctness, navigation, and form validation gate every PR. This is the offline tier in references/production-smoke-testing.md, which has a starter spec.
2. Assert the browser console is clean in every flow. WASM panics, failed requests, and CSP blocks surface only as console or network errors, so a UI that looks correct in a screenshot can still be panicking on every interaction.
3. Once a local node is running, re-run the same flows against the gateway-served webapp (the iso tier). Reaching your app there needs frameLocator and an absolute-URL goto, because the gateway wraps every webapp in an iframe shell.

For interactive debugging rather than scripted specs, the Playwright MCP browser tools drive a running dx serve or local node directly. See the local-dev skill, "Debugging with Playwright".

References:

• references/ui-patterns.md - WebSocket connection models, gateway CSP, framework-specific patterns.
• references/production-smoke-testing.md - the four test tiers, the development-loop browser-validation recipe, and the iframe-shell Playwright idioms.

Phase 4: Build, Test, and Deploy

Set up the build system, CI, and deployment pipeline.

Implementation steps:

1. Set up build orchestration — either Makefile.toml (cargo-make) or plain Makefile

2. Add a preflight task that runs fmt, clippy, tests, and migration checks before publish

3. Add GitHub Actions CI workflow (runs on push and PRs)

4. Back up contract state to the delegate for network resilience

5. Add a production-liveness smoke test. A ~50-line Playwright spec asserting the gateway-hosted webapp mounts, vendored CSS loaded, and the browser console is clean catches CSP blocks, iframe-shell mistakes, and broken archives that no unit test reaches. See references/production-smoke-testing.md.

6. Check the gateway port and (optionally) tar reproducibility. The gateway runs on 7509 — older docs and scripts still reference 50509. For byte-reproducible webapp archives across build hosts, invoke tar with the GNU flags listed under "Tooling Preflight" in references/build-system.md.

7. If you'll ship more than one release, plan for a facade contract from day one. Without it, every release rotates the gateway URL users have bookmarked. The facade is a stable-URL indirection: a never-rebuilt contract whose state holds a signed pointer to the current web-container. See references/facade-pattern.md. Cheaper to design in now than retrofit later — retrofitting means asking every existing user to update their bookmark.

8. Plan contract-WASM stability before the first release. A cargo update in the workspace root must not silently rotate contract IDs. See references/build-system.md → "Per-contract lockfile isolation".

9. Test the upgrade path and make migration resumable. The dangerous inputs are old-state -> new-code and interrupted migration, neither exercised by testing the new version on fresh state. Add an old-format-load test and an interrupted-migration-recovery test, and make migration idempotent + resumable (in-progress marker cleared only on full success) + non-destructive + regression-gated + observable. See references/upgrade-and-migration.md.

References:

• references/build-system.md — build, CI, packaging, tooling preflight, per-contract lockfile isolation, contract-ID reproducibility caveat, pre-commit hook for stray .wasm.
• references/production-smoke-testing.md — iframe shell architecture, Playwright recipe for post-publish liveness checks.
• references/facade-pattern.md — stable-URL facade contract architecture for projects that ship more than one release.
• references/upgrade-and-migration.md — operational discipline for safe contract/delegate upgrades: the five migration properties (idempotent, resumable, non-destructive, regression-gated, observable), enumerating dynamic key families, the upgrade test harness, and staged reversible rollout.

Project Structure Templates

Dioxus (Rust) UI

my-dapp/
├── common/                    # Shared types between contract/delegate/UI
│   └── src/
│       ├── lib.rs
│       └── state/            # State definitions
├── contracts/                 # one subdirectory per contract crate
│   ├── room-contract/
│   │   ├── Cargo.toml
│   │   └── src/lib.rs        # ContractInterface implementation
│   └── profile-contract/      # add more as the app grows
│       └── ...
├── delegates/                 # one subdirectory per delegate crate
│   ├── chat-delegate/
│   │   ├── Cargo.toml
│   │   └── src/lib.rs        # DelegateInterface implementation
│   └── identity-delegate/     # add more as the app grows
│       └── ...
├── ui/
│   ├── Cargo.toml
│   ├── Dioxus.toml
│   └── src/
│       ├── main.rs
│       └── components/
├── Cargo.toml                # Workspace root
└── Makefile.toml             # cargo-make build tasks

TypeScript + Vite UI

my-dapp/
├── contracts/
│   └── my-contract/
│       ├── Cargo.toml
│       └── src/lib.rs        # ContractInterface implementation
├── delegates/
│   └── my-delegate/
│       ├── Cargo.toml
│       └── src/lib.rs        # DelegateInterface implementation
├── web/
│   ├── package.json
│   ├── vite.config.ts         # Injects contract/delegate keys at build time
│   ├── tsconfig.json
│   ├── index.html
│   └── src/
│       ├── index.ts           # Entry point, connection flow
│       ├── freenet-api.ts     # FreenetWsApi wrapper
│       ├── delegate-api.ts    # Delegate FlatBuffers message building
│       ├── identity.ts        # Identity management (delegate + fallback)
│       ├── types.ts           # Shared TypeScript types
│       └── components/        # UI components
├── Cargo.toml                 # Workspace root (contracts + delegates)
└── Makefile                   # Build orchestration

Reference Project

River demonstrates all patterns:

• Contracts: contracts/room-contract/
• Delegates: delegates/chat-delegate/
• UI: ui/
• Common types: common/

Key Dependencies

Track the versions River (the reference dApp) uses. Mismatched versions cause deserialization failures, missing features, and "variant index out of range" errors. Check River's workspace Cargo.toml before pinning.

As of May 2026 — River pins freenet-stdlib = "0.6.0" but the upstream crate is now 0.8 (0.6 → 0.7 added Base58-stringified contract_states keys in NodeDiagnosticsResponse; 0.7 → 0.8 hardened wire-boundary enums with #[non_exhaustive] and removed the world-known DEFAULT_CIPHER / DEFAULT_NONCE constants). If you build only against River, mirror its pin; if your code links into stdlib 0.8 directly, you need the bumped version and the wildcard match arms / random cipher generation documented in references/delegate-patterns.md.

# Workspace-wide (Cargo.toml) — track this against stdlib 0.8 once River bumps.
freenet-stdlib = { version = "0.8", features = ["contract"] }
freenet-scaffold = "0.2.2"
freenet-scaffold-macro = "0.2.2"

# UI crate (ui/Cargo.toml): enables WebApi/WebSocket helpers
freenet-stdlib = { workspace = true, features = ["net"] }

# UI framework
dioxus = { version = "0.7.3", features = ["web"] }

The contract feature is required for contract crates targeting wasm32-unknown-unknown; use the delegate feature for delegate crates. The net feature pulls in WebApi for the UI.

TypeScript UI

For UIs built with TypeScript + Vite (Option B in Phase 3), depend on the matching @freenetorg/freenet-stdlib release:

{
  "dependencies": {
    "@freenetorg/freenet-stdlib": "^0.2.0"
  },
  "devDependencies": {
    "vite": "^6.0",
    "typescript": "^5.0",
    "sass": "^1.0"
  }
}

The TS package v0.2.0 brought the API to parity with the Rust client: FreenetWsApi with promise-based get/put/update/subscribe/ disconnect (await api.X(...)), full ResponseHandler including onContractNotFound/onSubscribeResponse/onClose, inbound ReassemblyBuffer, and transparent outbound chunking for payloads

512 KB. Callbacks still fire alongside promises for backward compatibility; the default request timeout is 30 s. See references/ui-patterns.md for the full pattern and a warning about the private sendRequest cast used for delegate messages until a public builder lands.

Security: removed encryption defaults

stdlib v0.6.0 (PR #75) removed the public constants DEFAULT_CIPHER and DEFAULT_NONCE to close a CVE-class issue (world-known keys leaked into any binary that imported them). Delegates that previously used these must now generate random values per session — e.g. let key: [u8; 32] = rand::random(); let nonce: [u8; 24] = rand::random();. Code still referencing the old constants will fail to compile against stdlib 0.6 or newer.

Improving This Skill

This skill is designed to be self-improving. When encountering issues while using this skill, agents should file GitHub issues or submit PRs to improve it.

When to File an Issue

File an issue at freenet/freenet-agent-skills when:

• Instructions are unclear or ambiguous
• Information is missing for a common use case
• Code examples don't compile or are outdated
• Patterns don't match current River implementation
• A referenced API has changed

How to File an Issue

gh issue create --repo freenet/freenet-agent-skills \
  --title "dapp-builder: <brief description>" \
  --body "## Problem
<describe what was unclear or incorrect>

## Context
<what were you trying to accomplish>

## Suggested Improvement
<optional: how the skill could be improved>"

Submitting a PR

For concrete improvements:

# Clone and create branch
gh repo clone freenet/freenet-agent-skills
cd freenet-agent-skills
git checkout -b improve-<topic>

# Make changes to dapp-builder/SKILL.md or references/*.md
# ... edit files ...

# Submit PR
git add -A && git commit -m "dapp-builder: <description>"
gh pr create --title "dapp-builder: <description>" \
  --body "## Changes
<describe improvements>

## Reason
<why this helps>"

What Makes a Good Improvement

• Fixes factual errors or outdated information
• Adds missing patterns discovered while building a dApp
• Clarifies confusing instructions based on real usage
• Adds test examples that would have helped
• Updates code to match current Freenet/River APIs