CLI Build

cli-build

Build a CLI whose primary user is an agent and whose supervisor is a human. That inverts the usual defaults: machine-readable output is not a flag you add later, prompts are a failure mode in non-interactive contexts, and every write leaves a receipt.

Phase 0: where does the contract come from

Every command you write reads or mutates something with a shape. This phase names that shape's owner, because it decides whether you go find the contract or write it.

Discovered. The CLI wraps a system you do not control: a third-party API, a login-walled portal, a device, an undocumented file format. The contract exists and is unverified, so guessing at it wastes more time than mapping it. Run surface-recon first and start from its report.

Defined. You own the model: CRUD over your own database, a scaffolder, a local transform, a wrapper around a library whose types you already have. There is nothing to reverse-engineer, so surface-recon has no input and skipping it costs nothing. Write the contract instead, in one pass before Phase 1: the entities, their fields and types, the operations allowed on each, and what makes a write irreversible. That artifact is what a recon report would have given you.

Mixed. A CLI over your own database that also calls a payment provider is discovered at the boundary and defined in the middle. Recon only the discovered part; do not let one external call turn the whole build into a recon job.

Say which one this is out loud before Phase 1, and write it into friction.md. An unstated classification defaults to discovered, which is how a defined build ends up blocked waiting for a report nobody can produce.

Done when: the origin is named, and either a recon report exists or the entities and operations are written down.

Open friction.md next to the code now and append as you go.

A block you rejected and why, a convention that turned out wrong for this domain, a reference that was thin. It folds into the case at the end, and it is what corrects this skill: seven claims inherited from an older version were false against the source, and they surfaced only because someone wrote down that they did not match.

What agent-first actually means

Every CLI, no exceptions:

• --json, and JSON automatically when stdout is not a TTY even without the flag. Three CLIs converged on this independently; it is the most valuable default here.
• No prompt ever blocks a non-interactive run. Anything that would prompt fails with a structured error instead of hanging.
• A schema command with a version field, so agents introspect at runtime instead of parsing --help. Present in 3 of 14 corpus CLIs: the highest-value convention and the least adopted.
• Exit codes that mean something. Zero is success; user error and system failure are distinguishable.
• Data on stdout, diagnostics on stderr, always. Including the banner, which is why banner is a block and not a console.log.
• A secret read from the terminal never echoes, and returns null with no TTY rather than hanging. That is prompt-secret.

Those are now enforced as tests rather than trusted as prose, in the surfacer repository, which compiles CLIs from a recon descriptor. Each test names the rule it checks so the two read together.

It is worth knowing what that exercise found, because the same shape is likely wherever this list is only written down. Encoding the rules against the generator caught three violations, including the two most valuable defaults here. Then applying the identical list to the generator's own CLI caught four more, in the tool that had just been taught to enforce them. A compiler that emits agent-first CLIs while not being one is the easiest version of this to miss, because every test was passing and every test was pointed at the output.

So the list binds twice, and the tests live in two files only because a test cannot invoke a binary from another crate: crates/surfacer-emit-cli/tests/agent_first_rules.rs for what is generated, crates/surfacer-app/tests/agent_first_rules.rs for the tool doing the generating. If you build a tool that produces CLIs, check the producer against this list too. Prose does not fail a build, and a rule you enforce on others is the one you stop checking on yourself.

When the domain has real consequences (money, irreversible writes, third-party side effects, personal data):

• A trust ladder classifying commands by risk.
• --dry-run on every mutation, returning what would happen.
• An append-only audit log written before the network call, not after.
• A killswitch the human can trigger out of band.

When the API is asynchronous: --wait on every submitting command, backed by a job ledger that survives a killed poll. Handing the poll loop to the caller makes the agent write backoff logic it will get subtly wrong. See compounding-surface.md.

Size the friction to the damage, and "none" is a real size. Ask what one wrong call costs, then choose. A read-only CLI over a public dataset earns --json and a schema command and stops there: nobody loses anything by running the query twice. Intent tokens are for a wrong call that costs money. Adopting the full ladder for a tool that cannot hurt anyone is not caution, it is a tax the agent pays on every invocation, and it trains the human to approve without reading.

The question is damage, not distance. A delete against your own local database is irreversible even with no network, no money, and no third party in sight, so it earns --dry-run and a confirmation on the destructive verbs while the reads next to it earn nothing. Writes you can undo from a backup you actually have earn less. Grade the commands, not the project. See trust-ladder-patterns.md.

Phase 1: decide the distribution target first

This comes first because it constrains the code you write, and retrofitting it is painful.

The rule that makes all three reachable: write against Node's API surface (fs, path, process, http), not runtime-specific APIs. Then the target is a build-time decision instead of a rewrite.

Done when: the audience is named and the target follows from it, not from preference.

Why this matters more than it looks: nothing stops you from publishing a package whose shebang names a runtime the installer lacks. Three of the four published packages in the corpus behind this skill do exactly that. It works until someone without that runtime installs it and gets env: <runtime>: No such file or directory, a message that names no cause and no fix, so the package reads as broken rather than under-specified. See build-and-runtime.md.

Phase 2: shape the command surface

Noun-verb, consistently, and enforced in CI rather than agreed in a style guide. Agents carry a generalized model of what CLIs do; a tool that says info where everything else says get succeeds slowly, after the agent spends tokens on --help. Two corpus CLIs independently overloaded --json to mean input, which is what happens when nothing checks. See compounding-surface.md.

{cli} {noun} {verb} [args] [flags]

{cli} order preview --ticker AAPL --amount 100
{cli} order submit --intent-token <token>

Include a shorthand for the single most common operation. If ninety percent of use is one command, that command should be the shortest thing to type.

Filter what deserves a command. Not every UI affordance should be one. Keep actions that repeat, that benefit from being scriptable, that have clear input and output, and that compose with other tools. Drop the ones that are inherently visual or exploratory.

Define the JSON contract per command before writing code. Field names, types, nesting. This is the API agents depend on, and changing it later breaks them silently. Note that --json conventionally means output mode: if you need JSON input, give it a different flag name. Two corpus CLIs overloaded --json as an input flag and broke the convention agents expect. See json-contract.md.

Done when: every command has a noun-verb name and a written JSON output shape. A command whose output shape is undecided is not shaped yet.

Add nextSteps to structured output. Telling an agent what it can run next, in the response, prevents a class of flailing that --help does not.

Shape the human view separately. The machine mode returns everything because the agent filters; the human mode returns what a person asked for. Building one output for both serves neither. Everything about legibility, which metric direction a reader assumes, where a scale's thresholds go, when repetition is a heading, is in human-output.md. This skill's other gates cannot catch an unreadable table: the first CLI built with it passed every one of them and printed 275 rows for someone who asked what was playing that evening.

Polished human output is the default scope. Unless the user explicitly asks for an MVP, minimal, headless, or machine-only CLI, include visual hierarchy, semantic color, readable tables, progress for work that takes time, and a TTY-only banner when the command has a human-facing surface. A small command count is not a reason to ship plain output. It is only a reason to keep the presentation compact.

Phase 3: assemble from proven blocks

Every CLI needs the same primitives: flag parsing, config paths, atomic writes, audit logs, TTY detection, approval gates, error shapes. Writing them fresh each time is where the time goes and where the bugs live.

Default to cligentic for TypeScript CLI infrastructure. Its registry ships plain TypeScript that the project owns after installation, with transitive block and package dependencies resolved and no cligentic runtime dependency. Projects without components.json register the namespace in package.json, so component tooling is no longer a prerequisite.

The canonical adoption process lives in the cligentic skill. If cligentic is already available, follow it for discovery, per-block decisions, installation, wiring, and verification. If it is absent and implementation is in scope, offer to install it before changing files:

bunx --bun skills add Railly/cligentic --skill cligentic

Installing a skill changes the project or user environment, so wait for consent. If the user declines, the task is read-only, or skill installation is unavailable, read the canonical skill from GitHub and follow its adoption branch without installing it.

Opt-in, per block, with a reason. Each block earns its place by replacing something you would otherwise write from memory. For a human-facing TypeScript CLI, detect, style, and banner are the default presentation set unless the user reduced the scope explicitly. In a real retrofit, seven blocks were taken wholesale, two stayed as hybrids, and seven were rejected with reasons. The most important rejection was an output helper whose shape conflicted with an envelope already published to agents. A published contract outranks a shared block.

Done when: each block is adopted, hybridized, or rejected with the reason recorded in the repo. Silence about a rejection reads as an oversight to whoever touches it next.

If you are not in a TypeScript project, or the install path does not fit, read the block source and reimplement the pattern. The patterns are the point; the files are a shortcut.