Implement, scaffold, or modify an agent-ready CLI in a repository. Use when user says '/agent-ready-cli-build' or asks to build or implement a CLI from an existing spec, an OpenAPI file, or requirements. Produces a git-initialized repo on disk with source, tests, docs, and distribution instructions. For the full idea-to-delivery pipeline (story, spec, build, audit) prefer agent-ready-cli-end-to-end. Does not push, publish, or submit unless explicitly requested.
Installation
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Implement, scaffold, or modify an agent-ready CLI in a repository. Use when user says '/agent-ready-cli-build' or asks to build or implement a CLI from an existing spec, an OpenAPI file, or requirements. Produces a git-initialized repo on disk with source, tests, docs, and distribution instructions. For the full idea-to-delivery pipeline (story, spec, build, audit) prefer agent-ready-cli-end-to-end. Does not push, publish, or submit unless explicitly requested.
Use this skill to build or modify a CLI in a repo.
The primary deliverable is a git-ready repo on disk: source code, executable wiring, passing tests, docs, distribution instructions, local commits, and verification evidence — ready for the user to review and push. Pushing, publishing, opening PRs, or releasing are separate actions that require explicit user instruction.
Included References
references/agent-ready-cli-checklist-v2.md — canonical checklist and scoring rubric.
references/frameworks-and-implementation-guidance.md — framework selection and implementation guidance.
Scope and Routing
Triggering is defined in the frontmatter description. In scope:
building a new CLI end-to-end from a spec, OpenAPI file, or requirements;
adding agent-ready behavior to an existing CLI (--json, stdin, dry-run, auth status, exit codes, verification commands, tests);
producing a ready-to-test local package/binary with docs and a verification transcript.
For the full idea-to-delivery pipeline (story and spec included), route to agent-ready-cli-end-to-end.
Inputs
Input
Required
Notes
Contract source
Yes
A CLI spec (from agent-ready-cli-spec), an OpenAPI file (preferred when the CLI wraps an API), or plain requirements.
Target directory
Yes
Where the repo lives or should be created.
CLI name
No
Default: derive from spec/API title.
Language/framework
No
Default: choose per references/frameworks-and-implementation-guidance.md and say why.
Credentials for live testing
No
Env var names only — never ask the user to paste secrets. A credential-helper command the user points to (e.g., export GITHUB_TOKEN=$(gh auth token)) is also fine. Enables the optional live-endpoint step.
Distribution targets
No
npm, Homebrew, pipx/uvx, GitHub Releases, Docker. Default: the natural one for the chosen language, documented for the rest.
Any input can arrive as a path on disk, a URL, a GitHub repo, or text pasted into the conversation. Never go looking on disk and report what you found there.
Opening ask (only when the contract source or target directory is missing)
Ask once, like a person. Do not announce that a directory is empty, do not cite these instructions, do not explain what the skill requires — just ask:
Do you have a CLI spec, an OpenAPI file, or written requirements for this? Send them whichever way is easiest — a path on disk, a link, a GitHub repo, or just paste them here. And where should the repo live?
If you don't have anything written down, no problem — let's talk it through. What should this CLI do, and who's going to use it: you at a terminal, a coding agent, CI, or all three?
If they have no files, have the conversation — that is the input round. Cover the table above in plain sentences, a couple at a time. Never present it as a numbered questionnaire.
A placeholder API server URL is normal — never stop over one.example.com, localhost, {host}, TODO and the like are placeholders, not missing inputs. Wire the base URL through config (--base-url > TOOL_BASE_URL > config file > spec default), ask once in passing for the real one, and proceed either way — skipping only the live-endpoint check, logged under Assumptions.
Then run unattended — complete the build without pausing, and log defaults chosen under Assumptions in the final summary. If only an OpenAPI file or requirements were given (no spec), first produce a condensed command contract (per agent-ready-cli-spec's derivation rules) inside docs/cli-spec.md, then build against it.
Actual Deliverables
Default: git-ready repo on disk
Unless the user explicitly asks to push/publish/submit, deliver:
source code changed/created in the target repo;
git initialized (git init if not already a repo) with a .gitignore and meaningful local commits at milestones — scaffold, commands, tests, docs — so the history is reviewable; never git push;
an installable companion skill — SKILL.md at the repo root, in Agent Skills format, so it serves both discovery paths from one file: an agent already working in the repo reads it in place, and every other agent gets it via npx skills add <owner>/<repo>. A runbook nobody can install is invisible to an agent that is not already in your repo, which is most of them.
Frontmatter: name (lowercase, hyphens, matching the CLI), and a description that says what the CLI does and when to reach for it — the description is the only part an agent reads before deciding to load the skill.
Body: the four or five workflows that matter, as exact commands, not prose. How to confirm authentication before starting. Which commands mutate, and their dry-run form. The verification command for each workflow, because a zero exit code is not evidence the thing happened.
Not a second copy of the README. The README is for a human deciding whether to adopt the tool; the skill is for an agent already driving it. Duplicated prose costs the agent context and tells it nothing.
Record the install line in DISTRIBUTION.md next to npm publish, and check the skill loads (npx skills add <path-or-repo> --list) before declaring done.
an update check — tool update --check --json, reporting installed version, latest version, whether an update is available, and the exact update command. Implement it once a distribution channel is chosen. When nothing is published yet, still add the command and have it report that no channel is configured — then say so in DISTRIBUTION.md rather than omitting the command. Checklist category 11 scores the capability, not the registry.
DISTRIBUTION.md — exact, copy-paste instructions for every relevant channel: npm (npm publish steps, npx usage, semver/tagging), Homebrew (formula or tap steps), pipx/uvx, GitHub Releases (binary + checksums), Docker if relevant. Instructions only — do not run publish commands. Include the companion skill's one-line install (npx skills add <owner>/<repo>) alongside the package channels;
verification transcript saved to artifacts/agent-cli-eval.md;
concise summary of changed files, commands run, and assumptions made.
Node/npm deliverables
Depending on scope:
local npm package ready to test via npm link, npm test, npx . --help, or equivalent;
packed .tgz from npm pack, smoke-tested from tarball if requested;
publish-ready package only if requested;
published npm package only after explicit instruction and credential confirmation.
Python deliverables
local package with pyproject.toml console script and tests passing;
wheel/sdist built and smoke-tested in a clean venv if requested;
upload only after explicit instruction.
Go/Rust deliverables
source implementation with tests passing;
built binary under dist/ or equivalent if requested;
checksums/release artifacts only if requested.
Build Workflow
1. Confirm or create the implementation contract
Require or produce (in docs/cli-spec.md if absent):
command tree;
JSON schemas (reuse OpenAPI components/schemas when wrapping an API);
auth model (derive from OpenAPI securitySchemes: apiKey → env var; OAuth → token env var/file + auth login; none → no auth commands);
config precedence, including API base-URL config when wrapping an API;
dry-run/confirm behavior;
verification commands;
error and exit-code tables (default table in references/frameworks-and-implementation-guidance.md — don't invent a new scheme per project);
framework choice;
test plan.
Completion criterion: code can be judged against a concrete spec.
2. Choose framework by scope
Default guidance (details in references/frameworks-and-implementation-guidance.md):
oclif for serious multi-command Node/TypeScript product CLIs with plugins/generated docs.
commander/yargs/clipanion/cac for smaller Node CLIs.
Go/Rust when single-binary distribution and low runtime friction matter.
Python Typer/Click for Python/data-native workflows.
TUI/prompt frameworks only on top of headless commands.
Completion criterion: framework choice states why at least one plausible alternative was not chosen.
Completion criterion: core commands work without a TTY.
4. Add tests and docs
Tests must cover:
help/version;
JSON parseability;
stdout/stderr separation;
stdin/piped input where relevant;
no-TTY behavior;
error and exit codes;
dry-run/confirm;
verify-after-action workflow.
Network-dependent tests must not require live credentials: mock/stub the API by default so npm test/pytest passes in a clean clone. Preferred pattern: start a local HTTP stub server in the test suite and point the CLI's --base-url/TOOL_BASE_URL override at it — this exercises the real HTTP layer and justifies the base-URL config. (Note: interception libraries like nock do not catch Node's native fetch/undici by default.)
Docs must include: quickstart; install (matching DISTRIBUTION.md); top workflows as copy-paste commands; auth setup; examples; troubleshooting; safety model.
Completion criterion: tests pass and docs include copy-paste workflows.
5. Optional: live endpoint check, then agent eval
Live endpoint check (optional, never blocking): if the API's server URL is real and required credentials exist as env vars (check with test -n "$VAR" — never print values), verify the CLI works against the real API: run auth status, one read-only list/get command, and confirm parseable --json output. If credentials or network are missing, skip, use the mocked test evidence, and record the skip in the transcript.
Agent eval: run the loop with real commands in a safe environment:
Save the transcript with actual commands and outputs to artifacts/agent-cli-eval.md.
6. Self-check against the bundled checklist before declaring done
You are building the best tool for the people and agents who will use it. You are not building for the checklist. The checklist is a compass, not a destination — it exists because it usually points at what those users need, and it is wrong whenever it does not. Where following an item would make this CLI worse, do the better thing and record the deviation with its reason. Never add a flag nobody will use in order to satisfy a line: a --plain mode on a CLI with no tabular output is noise. Never drop a capability to dodge an item you might not satisfy perfectly — a CLI that removes stdin support has not improved, whatever any number says. The score is direction, never a target, and this skill is deliberately given no number to reach.
references/agent-ready-cli-checklist-v2.md is the rubric this build will be scored against. Walk it, category by category, before you report finished. For each item: satisfied (name the evidence — a command you ran, a file you wrote), deferred (say why, e.g. no distribution channel chosen yet), or genuinely not applicable (say why — never because it was inconvenient to satisfy).
Do not skip the items this skill's deliverables list does not spell out. The deliverables are a floor, not the rubric. Anything the checklist scores and the build can satisfy, the build should satisfy — and where it does not, the final summary must say so rather than leave an auditor to discover it.
Exercise the three --json parser-error paths here (unknown subcommand, unknown flag, missing required option). They are the most commonly missed item, they cannot be verified by reading, and a passing test suite does not imply they work — see references/frameworks-and-implementation-guidance.md.
Completion criterion: every checklist category has a verdict with evidence, and unmet items appear in the summary under Assumptions rather than being discovered later by an audit.
Completion criterion: final report cites actual commands and outputs, and states whether live API testing ran or was skipped.
Submission Boundary
Do not push, publish, open PRs, or release unless explicitly asked.
Deliverable ladder:
local repo ready to test;
local commits on a branch (default deliverable);
pushed branch;
opened PR;
packed artifact;
published package/release.
Levels 3–6 require explicit instruction. DISTRIBUTION.md documents how to do 5–6; the skill does not do them.
Output Format
End with this summary in the conversation:
# Build Summary: [tool]
Status: ready-to-test (committed locally, nothing pushed — the default) / pushed / PR opened / packed / published
Repo path: ...
Commits: [git log --oneline]
Tests: [command] → N passing (no credentials required)
Live API check: performed (evidence) / skipped (reason)
Changed files: ...
Assumptions: ...
Next actions: [exact commands to test locally; publish steps are in DISTRIBUTION.md]
Common Pitfalls
Snippet-only output. The deliverable is a repo state, not a pasted code sample.
Prompt/TUI-only implementation. Agents need headless commands.
JSON polluted by banners/spinners.--json stdout must parse directly.
No verification transcript. "Tests pass" is not the same as an agent workflow proof.
Publishing by surprise. Never push/publish/release without explicit permission.
Tests that need secrets. A clean clone must pass tests with no credentials set.
Uncommitted work. Milestone commits make the deliverable reviewable and recoverable; end with a clean git status.
Printing secrets. Check env vars with existence tests; never echo values into logs or transcripts.
Docs are updated; DISTRIBUTION.md covers every relevant channel with copy-paste steps.
An installable companion skill exists: root SKILL.md, agentskills frontmatter, workflows as exact commands, auth check, mutating commands marked with their dry-run form, a verification command per workflow. Distinct from the README. Its one-line install is in DISTRIBUTION.md and has been checked to resolve.
tool update --check --json exists, and reports honestly when no distribution channel is configured.
--help and --version work.
--json output parses — including for the framework's own parser errors: unknown subcommand, unknown flag, missing required option. Each emits a valid envelope and exits 2. Run all three; do not infer from a passing suite.
E_INTERNAL output carries version, and a report path where one exists.
Every checklist category has a verdict with evidence; unmet items are named in the summary.
stdout/stderr are separated.
Mutations have dry-run/confirm where applicable.
Verification command proves results.
Agent eval transcript is saved; live API check performed or skip recorded.
Final status says ready-to-test, ready-to-submit, or published with evidence, plus assumptions made.
