SKILL.md
SOTA CLI & Developer-Tool UX
Purpose
Expert-level rules for building and auditing command-line tools: the grammar of commands/flags/args, config layering, human-vs-machine output, TTY-aware interaction, exit codes, signal handling, lifecycle behavior, and distribution. The core thesis: a CLI has three users — a human at a TTY, a script in CI, and an AI agent driving the tool (a stricter script) — and every design decision must serve all three without flags-gymnastics. Rules are imperative with rationale and good/bad terminal examples; every rules file ends with an audit checklist. Load only the files relevant to the task via the index below.
BUILD mode
When designing or implementing a CLI:
- Design the command grammar before writing code. Write the
--helptext
first: subcommand tree, every flag (short + long), args, examples. If the help is hard to write, the grammar is wrong (rules/01).
- Defaults carry the common path. A new user must get a useful result with
zero flags. Anything required for the 80% case is a design bug (rules/01 §3).
- Split the streams from day one: primary output → stdout, everything else
(logs, progress, prompts, errors) → stderr. Retrofitting this breaks users' pipes (rules/02 §1).
- Make it pipe-safe by default: detect TTY; disable color/progress/prompts
when piped; honor NO_COLOR; ship --json with every listing/reading command from v0.1, because output format is API (rules/02).
- Treat exit codes, flags, env vars, and JSON shapes as a public API:
document them, version them, deprecate — never repurpose or remove without a cycle (rules/02 §3, rules/03 §7).
- Build the unhappy paths with the happy path: Ctrl-C cleanup,
--dry-run
on mutating commands, re-runnable/resumable operations, stdin-closed CI behavior, offline behavior (rules/03).
- Plan distribution early: single static binary if the ecosystem allows,
checksums + signatures, completions and man page generated from the same source as --help (rules/03 §8, rules/04).
- Before declaring done, run every relevant Audit checklist against your
own tool, including the brutal smoke test: tool cmd | cat, tool cmd > out.txt 2> err.txt, tool cmd < /dev/null, echo $?.
AUDIT mode
When reviewing an existing CLI:
- Identify the surface: parser setup (argparse/clap/cobra/etc.), main/entry
point, output and error paths, signal handlers, config loading, install/ release scripts. Load the matching rules files.
- Run the tool, don't just read it. Minimum probe set:
- tool --help, tool <sub> -h, tool --version, tool definitelynotacmd - tool list | cat and tool list | head -1 (color codes? broken pipe panic?) - tool list > /dev/null — does anything still reach the human? (it should, on stderr) - tool mutate < /dev/null — hang waiting for a prompt = CI killer - echo $? after success, after a usage error, after a real failure - Ctrl-C mid-operation: prompt state restored? partial state cleaned or resumable?
- Then verify in code what can't be probed: config precedence order, secret
handling, temp/state file locations, update/telemetry behavior.
Severity conventions
- Critical — corrupts data or destroys trust: destructive op with no
confirmation/--force and no dry-run; non-zero work reported as exit 0 (or vice versa) so CI lies; prompt hangs forever with stdin closed; secrets echoed to terminal/logs/argv; auto-update or telemetry without disclosure.
- High — breaks scripting or interrupts users: machine output polluted by
ANSI codes/log lines on stdout; no --json on listing commands; SIGINT leaves corrupt partial state; undocumented/colliding exit codes; config precedence nondeterministic; breaking flag removal without deprecation.
- Medium — erodes usability: missing long forms; required flags on common
path; no progress on >2s ops; errors without remediation; $HOME dotfile litter instead of XDG; no completions; no -q/-v; >500ms --help.
- Low — polish: help without examples, no suggest-on-typo, inconsistent
subcommand naming, missing man page, table borders in output.
Finding format
[SEVERITY] <one-line title>
Where: <file:line | command invocation that reproduces it>
Rule: <rules-file §section>
Issue: <what is wrong, with observed evidence (transcript or code)>
Impact: <who breaks: the human at the TTY, the script in CI, an agent driving the tool, or all>
Fix: <specific change; exact flag/stream/exit-code where load-bearing>
Order findings by severity; one finding per root cause; include the reproducing command line whenever the issue was observed by running the tool.
Rules index
| File | Read this when... |
|---|---|
rules/01-commands-flags-config.md |
Designing/auditing the command surface: subcommand grammar (noun-verb consistency), POSIX/GNU flag conventions, short/long forms, -- separator, args vs flags vs stdin, defaults, dangerous-op flags, config precedence chain, env var naming, help text quality, suggest-on-typo. |
rules/02-output-interaction.md |
Anything the tool prints or asks: stdout vs stderr contract, --json/--plain, TTY detection, color and NO_COLOR/TERM, exit codes as API, streaming vs buffering, progress indication, prompts and --yes/CI behavior, quiet/verbose levels, error message anatomy, --debug vs stack traces, agent-facing design (--fields, confirmation envelopes, frozen output contracts). |
rules/03-behavior-lifecycle.md |
Runtime behavior and tool lifecycle: startup latency, idempotent re-runnable commands, --dry-run, --self-test for tools whose output is a verdict, SIGINT/SIGTERM handling, crash-only resumable design, network-op responsiveness and offline behavior, self-update caution, telemetry consent, XDG base directories, shell completions, semantic versioning of the CLI surface. |
rules/04-distribution-docs.md |
Shipping and documenting: single-binary packaging, install channels (brew/curl-script/registry) with checksums + signatures, docs generated from the --help source of truth, man pages, README quickstart, version pinning for CI. |
Top-10 non-negotiables
- Primary output to stdout, everything else to stderr — logs, progress,
prompts, warnings, errors. tool x | next must receive only data. (rules/02 §1)
- Pipe-safe by default: TTY-detect; no color, no spinners, no prompts when
piped; honor NO_COLOR (set and non-empty) and TERM=dumb; --json on every command that lists or reads data. (rules/02 §2, §4)
- Exit 0 only on success; distinct, documented nonzero codes for usage
error vs operational failure; never exit 0 after printing an error. (rules/02 §3)
- Every flag has a long form; short forms only for the frequent few;
GNU/POSIX syntax (-f, --flag, --flag=value, -- ends flags). (rules/01 §2)
- No required flags on the common path — good defaults; and the reverse:
destructive operations require explicit --force/typed confirmation on a TTY and refuse (don't hang) without one. (rules/01 §3, §5)
- Every prompt is skippable:
--yes/--no-inputequivalent exists; stdin
closed or non-TTY ⇒ use default or fail fast with the flag to pass — never block CI. (rules/02 §6)
- Errors say what failed, why, and what to do next — path, value, and the
exact command to run; stack traces only under --debug. (rules/02 §7)
- Ctrl-C is sacred: first SIGINT exits promptly with cleanup and exit code
130, second SIGINT force-quits; terminal state always restored; interrupted work is resumable or rolled back. (rules/03 §3)
--dry-runon every mutating command, printing the real plan through
the real code path. (rules/03 §2)
- The CLI surface is a semver'd API: flags, exit codes, JSON fields, env
vars. Deprecate with warnings for a full cycle; never silently change meaning. Config precedence is exactly flags > env > project config > user config > defaults, documented. (rules/01 §6, rules/03 §7)