Convention

Iceberg Convention

What this skill does

The Iceberg Convention is a coding convention for senior and staff engineers whose goal is to make the code juniors (and AI agents) actually touch maximally frictionless. It does this by quarantining complexity behind an architectural airgap: seniors absorb the sophistication (types, state machines, custom linters, architecture tests, observability) below the waterline, and the surface above it reads as linear, synchronous-looking, imperative business logic.

This skill helps you apply the convention while writing or reviewing code. It is organized around three modes:

• Authoring — you are about to write or modify code. Apply the rules during generation.
• Audit — you are reviewing an existing codebase against the convention. Produce a structured report.
• Bootstrap — you are setting up the convention's enforcement scaffolding in a fresh project.

This is a language-agnostic convention

The rules describe patterns (airgap separation, branded types, discriminated unions, exhaustive matching, pure cores, runtime tracing, written rationale). The patterns have implementations in every mainstream typed language; their syntax differs and their native support varies, but the intent is invariant.

When applying the convention:

1. Detect the target language from the repo — package.json, pom.xml, Cargo.toml, pyproject.toml, *.csproj, go.mod, mix.exs, Gemfile, composer.json, etc. If multi-language monorepo, identify the language of the file you're about to write or review.
2. Translate each rule's pattern into the language's native idiom. The enforcer category (below) tells you what kind of tool; the idiomatic-tool lookup is governed by the WebSearch / WebFetch directive in the paragraph immediately below the numbered list.
3. Be honest when a language's native support is weak. Some rules enforce more strictly in Rust or Kotlin than in Python or Go. Flag the gap explicitly; do not pretend a type UserID string in Go is as safe as a Rust newtype struct.

Do not hard-code tool names into your output. The rule is "branded types with a validation constructor"; the enforcer is "whatever this language calls an AST-level linter, plus whatever this language calls nominal typing."

When the idiomatic enforcer for a (language, Category A–F) pair is unknown:

• If WebSearch is available: MUST invoke WebSearch with query: "<language> <category-name> linter OR architecture test idiomatic 2025" to verify the tool is currently maintained.
• Else if WebFetch is available: MUST invoke WebFetch on the ecosystem's curated tool-list page (e.g., https://awesome-rust.com, https://github.com/ziadoz/awesome-php, https://github.com/vinta/awesome-python).
• If neither is available: MUST state in the response that the enforcer recommendation cannot be verified, and flag it as Unverified — review before installing.

Do NOT recommend a tool you have not verified is currently maintained.

The convention's Axiom of Enforcement governs everything: if a rule is not enforced by the compiler, a linter, an architecture test, or a build gate, it does not exist. Your output is weighted toward generating enforcement (configuration, custom rules, type scaffolding, ADRs), not prose advice.

Stack detection recipe (shared by all modes)

Before applying any rule, MUST detect the target language(s). Invoke Glob in a single parallel batch with patterns:

• **/package.json · **/pom.xml · **/build.gradle · **/build.gradle.kts
• **/Cargo.toml · **/pyproject.toml · **/requirements.txt · **/setup.py
• **/*.csproj · **/*.fsproj · **/go.mod · **/Gemfile
• **/mix.exs · **/composer.json · **/Package.swift

For every hit, MUST invoke Read with limit: 80. Record detected languages + manifest paths in one line before proceeding.

If no manifest matches, MUST invoke AskUserQuestion:

AskUserQuestion
  questions: [{
    question: "No recognized manifest found. Which language family is this code?",
    header: "Language",
    multiSelect: false,
    options: [
      { label: "TS/JS/Python", description: "TypeScript, JavaScript (Node/Deno/Bun/browser), or any Python ecosystem." },
      { label: "JVM / .NET",   description: "Java, Kotlin, Scala, Clojure; or C#, F#, VB.NET." },
      { label: "Rust / Go",    description: "Cargo-based or embedded Rust; or standard Go toolchain." },
      { label: "Other",        description: "Ruby, Swift, Elixir, PHP, etc. — describe in chat; convention may have reduced enforcement." }
    ]
  }]

Step 1 — Select the mode

Skip AskUserQuestion ONLY if the user's request contains one of the keyword sets below:

• Authoring: "implement", "write", "add", "refactor", "fix", "build [feature]"
• Audit: "audit", "review the codebase", "check the codebase", "how well does this follow"
• Bootstrap: "bootstrap", "set up", "scaffold", "configure conventions", "new project"
• Education: "what would the convention say", "explain the convention", "how does the convention handle"

If the request matches keywords from ≥2 rows, OR matches no row cleanly, MUST invoke AskUserQuestion:

AskUserQuestion
  questions: [{
    question: "Which Iceberg Convention mode applies to this task?",
    header: "Mode",
    multiSelect: false,
    options: [
      { label: "Authoring", description: "Writing or modifying code under the convention." },
      { label: "Audit",     description: "Reviewing an existing codebase; produces a findings report." },
      { label: "Bootstrap", description: "Setting up enforcement scaffolding (greenfield or retrofit)." },
      { label: "Education", description: "Read-only — explain what the convention would say about a design." }
    ]
  }]

After mode is determined, MUST invoke Read on the mode file:

The design-review-only sub-branch ("is this design compliant?") falls under Authoring — references/mode-authoring.md § Design Review.

Rule index (the 20 normative rules)

Cite rules by number. The full rationale and escape hatches for each rule live in references/rules.md — load it only when you need rationale (e.g. responding to user pushback), an escape hatch, or the language-native-support matrix for a specific rule.

Pillar 1 — API Airgap (structural boundary between business logic and infrastructure):

• §1.1 — tip is imperative and synchronous-looking (no raw futures/promises/observables/channels in tip-layer signatures) — enforcer B
• §1.2 — no infrastructure imports (ORMs, HTTP clients, loggers, tracers, cache, queues, feature flags) above the waterline — enforcer A
• §1.3 — functional core, imperative shell (business decisions are pure; I/O/time/randomness/mutation live in the shell) — enforcer D (primary), A (secondary)
• §1.4 — one file is enough for a typical change (if a change requires hopping 3+ abstraction layers, the abstraction is misplaced) — enforcer Design Review

Pillar 2 — Compiler-Driven Mentorship (mentorship belongs in the IDE, not the PR):

• §2.1 — no convention survives without an enforcer (unenforced rules train juniors to distrust the guide) — enforcer F (meta-audit)
• §2.2 — lint errors explain themselves (rationale + correct pattern + ADR link in every custom-rule message) — enforcer B (meta-lint on the custom-rules package)
• §2.3 — ban cognitive hazards at the AST level: accumulator reductions, nested ternaries, boolean flag params (doThing(true)), implicit timezone/locale/currency, silent catches — enforcer B
• §2.4 — every custom lint or arch rule ships with passing + failing snapshot tests — enforcer F (CI)

Pillar 3 — Defensive Type Engineering (encode domain constraints in types; IDE autocomplete mentors the junior):

• §3.1 — no raw scalars in domain signatures (every ID, money, duration, percentage, email, URL, enum-like string is a branded/nominal/opaque type) — enforcer C (primary), B (secondary)
• §3.2 — illegal states unrepresentable (discriminated unions / sum types / sealed hierarchies, not records of optional fields) — enforcer C (primary), B (secondary)
• §3.3 — validation boundaries are singular and explicit (one named constructor per branded type; direct coercion banned in tip) — enforcer B (primary), C (secondary)
• §3.4 — exhaustiveness is a compile-time obligation (every match over a domain union has a compile-fail sink for unhandled variants) — enforcer C

Pillar 4 — State-Machine Dictatorships (no ad-hoc boolean tuples encoding async lifecycles):

• §4.1 — no flag tuples (2+ booleans derived from the same underlying process = unadmitted FSM) — enforcer B (primary), Design Review (secondary)
• §4.2 — transitions are named, centralized, pure (reducer (state, event) → state or declarative machine; scattered mutations banned) — enforcer B (primary), A (secondary)
• §4.3 — happy path is a subset: FSMs enumerate idle + in-flight + success + recoverable-failure + terminal-failure; cancellation is a state, not an afterthought — enforcer Design Review (partial automation via variant count)
• §4.4 — UI renders from state (discriminator match), not from derived booleans passed across component boundaries — enforcer B

Pillar 5 — Observability as Documentation (runtime behavior doesn't decay; static docs do):

• §5.1 — every cross-boundary call (tip → berg, or service → service) emits a span, wrapped at the adapter not at call sites — enforcer E
• §5.2 — the "why" lives in ADRs, not comments (comments rot with their line; ADRs rot with their decision) — enforcer F (PR references ADR number)
• §5.3 — comment smells (restating code) are review-blocking; "why" comments (bug references, unobvious invariants, perf traps) are welcome — enforcer Design Review, B partial
• §5.4 — logs are for operators; traces are for readers. print("entering handler") is a junior-authored trace — delete it, instrument properly — enforcer B

Enforcer categories (A–F)

Every rule above maps to one or more of these. When proposing enforcement, identify the category, then pick the idiomatic tool in the target stack. Per-language canonical tool lists and the full tool-selection discipline live in references/enforcement-patterns.md — MUST invoke Read on it before proposing a tool you haven't used before in that ecosystem, or when authoring a project-specific custom rule (§2.2 discipline). The A–F definitions below are the routing surface; the catalogs are in the reference.

• A — Module-boundary enforcer. Fails the build when code in one layer imports from another it shouldn't. Enforces §1.2, §1.3 (via dep graph restriction), §4.2.
• B — AST-level linter. Fails on syntactic patterns. Enforces §1.1, §2.3, §3.1 (paired with C), §3.3, §4.1, §4.2, §4.4, §5.4.
• C — Type-level enforcement. Makes invalid states, values, or missing cases unrepresentable. Enforces §3.1, §3.2, §3.4. Native strength varies wildly — strong in Rust/Kotlin/Scala/Swift/F#/OCaml/Haskell/Elm; library-enabled in TS/C#/Java17+; check-time only in Python; weak in Go and JS-without-TS. Consult references/enforcement-patterns.md § Category C for the per-language matrix before committing.
• D — Purity enforcer. Pure modules have no I/O, time, random, network, DB, logger, env, or mutable global in their surface. Enforces §1.3. Three implementation approaches: (1) module-boundary restriction via Category A tools; (2) AST linter over @pure-tagged files via Category B; (3) language-level effect tracking (Haskell IO, PureScript Effect).
• E — Tracing instrumentation. Wraps every cross-airgap call in a distributed-tracing span with stable naming and required attributes. Enforces §5.1. Universal tool: OpenTelemetry, with language-specific SDKs and auto-instrumentation for common HTTP/DB/queue libraries. Wrap at the adapter, never at call sites — if juniors have to remember to add spans, they won't.
• F — PR-level check. Runs on every PR, gates merge on repo-level invariants. Enforces §2.1 (meta-audit), §2.4, §5.2. Always use the project's existing CI platform — installing a new one for this alone is a cognitive-load violation.

Honest gaps to remember: Go has no sum types (§3.2 becomes interface + known impls + review; weaker). Python has no runtime brands (§3.1 is check-time only; compensate with boundary validation via Pydantic/attrs + property tests). Dynamic languages (no static types) cannot enforce §3.1/§3.2 at the type level at all — lean on B + D. When you hit a gap, name (a) the rule, (b) the language limitation, (c) the closest available enforcer, (d) the residual risk review must cover.

When to load each reference

For assets/templates/ files (ADR skeleton, ADR-0001 example, CLAUDE.md fragment, PR template): MUST invoke Read on ${CLAUDE_SKILL_DIR}/assets/templates/<file>.md only in Bootstrap mode (or when explicitly adapting a template into a repo). Apply placeholder substitutions, then MUST invoke Write to the project's conventional destination path.

Hard invariants for every mode

Regardless of mode, these apply to anything you produce:

1. Never weaken a rule to fit existing code. If the current code violates the convention, say so. Do not pretend it complies.
2. Every rule you invoke cites its number. "This violates §3.1 (no raw scalars in domain signatures)" beats "this should be branded."
3. Every prescription names its enforcer category. If you say "use branded IDs," name which A–F category enforces it. If a specific tool is idiomatic for the target language, name it. If you cannot name an enforcer category, the prescription is folklore — rewrite it.
4. Distinguish berg from tip in every artifact. When you write code, mark it with the project's layering convention (path placement or a layer pragma). Enforcement rules cannot target what isn't labeled.
5. Treat AI agents as a first-class consumer of the surface. If you are writing a new API or enforcer, ask: does an autonomous agent producing code against this surface have the same or better guardrails as a human junior? If not, strengthen.
6. Never produce a visual diagram, chart, or widget for this task. This skill's deliverables are code, configuration, and structured text. Diagrams would be cognitive-load injection, which is what the convention is designed to prevent.
7. Pseudocode teaches the pattern; committed language implements it. The pseudocode in rules.md teaches the shape; when you generate actual code, use the target language's idioms — do not copy pseudocode verbatim.
8. Tool-invocation enforcement for user-visible actions. Every user-visible action step in this skill's mode files (ask-user, scan-repo, install-tool, write-file, run-build, gate-plan, and — only where governed — dispatch-subagent) MUST fire as a named tool invocation with full UX-critical parameters. This top-level index file itself dispatches no subagents; concrete Agent / EnterPlanMode / Bash invocations live in the mode files, governed by this invariant. Pure reasoning steps (classify-change, interpret-findings) are exempt. Prose-only action steps are folklore — the same failure mode §2.1 warns against, applied to this skill itself.

Default output expectations

• Authoring deliverables: the requested code in the target language, in the correct layer (tip vs. berg), plus the enforcement additions (custom lint rule, module-boundary config, ADR stub) the change necessitates — in whatever tool the target ecosystem uses.
• Audit deliverables: a Markdown report structured as in references/mode-audit.md, grouping findings by pillar, each finding citing its rule number, severity, concrete remediation, and the enforcer category (plus a specific tool recommendation if one is idiomatic for the target language).
• Bootstrap deliverables: configuration files appropriate to the stack, an ADR directory with ADR-0001 explaining adoption, a filled CLAUDE.md fragment, and a PR template.

When the user pushes back

The convention is demanding. Users may argue that a specific rule is overkill for their context. Respond as follows:

• If their argument is cognitive-load-based ("this makes the code harder to read for everyone"), take it seriously — the convention exists to reduce cognitive load; a rule that increases it in a specific case may deserve exemption. Document the exemption as an ADR.
• If their argument is effort-based ("setting up branded types is too much work"), the rule stands. The convention's entire premise is that seniors trade effort-below-the-waterline for frictionlessness-above. Name the tradeoff explicitly.
• If their argument is "we don't have juniors," redirect to the AI-agent rationale. Autonomous agents are the new juniors and their error modes are worse.
• If their argument is "this language doesn't really support that pattern," it is probably true and probably matters. Check the language-native-support matrix in references/rules.md for that specific rule. Acknowledge the gap, propose the closest available enforcer, note the residual risk that must be covered by review.
• If they insist on violating a rule, produce the code they asked for but include a clearly-marked comment (in the language's comment syntax) indicating CONVENTION: violates §X.Y — see ADR-NNNN and open an ADR stub documenting the exemption.

Additional resources

• For the 20 normative rules with full rationale, escape hatches, and the language-native-support matrix, see references/rules.md
• For per-language tool catalogs (A–F), tool-selection discipline, and the full rule→enforcer-category mapping, see references/enforcement-patterns.md
• For anti-pattern recognition — load in Audit mode and in any refactor task — see references/anti-patterns.md
• For Authoring-mode step sequence and the Design Review sub-branch, see references/mode-authoring.md
• For Audit-mode step sequence and findings-report structure, see references/mode-audit.md
• For Bootstrap-mode step sequence and scaffold-generation flow, see references/mode-bootstrap.md
• For the Bootstrap-mode ADR skeleton template, see assets/templates/adr.md
• For the Bootstrap-mode ADR-0001 adoption example, see assets/templates/ADR-0001-adopt-iceberg-convention.md
• For the Bootstrap-mode project CLAUDE.md fragment, see assets/templates/CLAUDE.md-fragment.md
• For the Bootstrap-mode pull-request template, see assets/templates/PULL_REQUEST_TEMPLATE.md

Now, per the §"When to load each reference" table above, MUST invoke Read on the mode file matching the selected mode, and proceed.