Script writing

<!-- cloud_safe: degrade -->

script-writing

When to use

• Creating a new TypeScript script in src/scripts/{name}.ts (linters, checks, generators, measure tools — run via ./scripts-run src/scripts/{name})
• Editing an existing script that prints progress or success lines
• Wiring a script into Taskfile.yml or taskfiles/*.yml
• Reviewing a PR that adds scripts and asking "why does this still print on minimal?"

Do NOT use this skill when:

• The content is a one-off / archival under scripts/ai_council/one_off_archive/ — those carry an _one_off_ prefix and are exempt from the verbosity convention
• The content is a shell entrypoint with secret prompts (install-keys, release confirms) → see § 3 Iron-Law carve-outs
• The content is a .mjs / Node script under scripts/cost/ — different runtime; convention covered in agents/settings/contexts/cost-tracking.md

Script vs other writers — critical test

Intent	Artifact
"Maintenance script the agent or CI runs"	This skill
"User types /foo to invoke"	command-writing
"Constraint the agent must always honor"	rule-writing
"Reference knowledge agents cite"	guideline-writing

Scripts orchestrate file checks, generators, and validators. They are neither user-invoked nor agent-routed — task and CI call them.

Procedure

0. Run the Drafting Protocol

Creating or materially rewriting a script that joins the linter / check family must go through Understand → Research → Draft from the artifact-drafting-protocol rule.

• Understand — what failure does this script catch that no existing check catches? Is it a one-off (then archive it) or evergreen?
• Research — ls src/scripts/check_*.ts src/scripts/lint_*.ts, grep for overlap, skim 1–2 peer scripts (e.g. lint_handoffs.ts, check_md_language.ts).
• Draft — propose name + one-line purpose first. Only fill the body after the shape is confirmed.

1. --quiet flag — argv check

Every check_*.ts / lint_*.ts script MUST accept --quiet so the silent Taskfile layer (§ 4) can suppress success-only output.

Canonical pattern (see e.g. lint_handoffs.ts):

const QUIET = process.argv.slice(2).includes('--quiet');
// ...
if (!QUIET) {
    console.log('✅ All clean');
}

Failure output (❌, non-zero exit) is never gated — failures must always print regardless of --quiet.

2. _lib/script_output helpers — preferred for new scripts

For anything richer than a single ✅/❌, import the verbosity-aware router instead of raw print():

import { info, success, warn, error, flush_summary } from './_lib/script_output.js';

info('Loading manifest');          // drops on silent + minimal
success('Wrote 3 files');          // collected at minimal, printed at verbose
warn('Skipping stale entry');      // stderr unless silent
error('Manifest missing');         // stderr always

flush_summary('Done — 3 entries'); // one-line summary at minimal

Resolution order (first wins, cached for the process):

1. AGENT_SCRIPT_VERBOSITY env (silent / minimal / verbose)
2. SCRIPT_OUTPUT_VERBOSE=1 alias (== verbose)
3. .agent-settings.yml → verbosity.script_output
4. Default minimal

The resolved level is exported back into AGENT_SCRIPT_VERBOSITY so child processes inherit it. Tests reset via reset_level() from the same module — see tests/lib/script_output.test.ts.

3. Iron-Law carve-outs — never silenced

The following surfaces MUST use plain print() and never the helpers, so verbosity settings cannot suppress them:

• Release confirms — every task in taskfiles/release.yml
• Secret prompts — install-anthropic-key, install-openai-key, setup-evals, install-hooks interactive sections
• runtime-e2e and test-triggers-live
• CI orchestration sentinels — _ci-start, _ci-end, root ci
• Any prompt that asks the user for confirmation per non-destructive-by-default — Hard Floor cannot be silenced

If unsure, check src/scripts/ai_council/one_off_archive/2026-05/README.md for the archived carve-out inventory.

4. Taskfile wiring — silent: true + {{.QUIET_FLAG}}

Every Taskfile task that wraps a --quiet-aware script MUST set silent: true and pass {{.QUIET_FLAG}} to the script:

# Per-task in taskfiles/*.yml:
tasks:
  lint-handoffs:
    silent: true
    cmd: ./scripts-run src/scripts/lint_handoffs {{.QUIET_FLAG}}

The QUIET_FLAG var is defined once at the root of Taskfile.yml and resolves to "" only when AGENT_SCRIPT_VERBOSITY=verbose:

# Root Taskfile.yml — already in place, do not duplicate:
vars:
  QUIET_FLAG:
    sh: '[ "$AGENT_SCRIPT_VERBOSITY" = "verbose" ] && echo "" || echo "--quiet"'

Carve-out tasks (release, install secrets, CI orchestration — see § 3) do not add silent: true and do not use {{.QUIET_FLAG}}.

5. Validate

• Run ./scripts-run src/scripts/skill_linter src/skills/script-writing/SKILL.md → 0 FAIL
• Run ./scripts-run src/scripts/{your-script} --quiet and the verbose path — exit code 0 on clean, non-zero on failure regardless of flag
• If the script uses _lib/script_output, add a test under tests/ patterned on tests/lib/script_output.test.ts — assert silent / minimal / verbose behave per § 2
• Run the full CI pipeline locally (see Taskfile.yml in this repo for the script list) — must exit 0 except for tolerated warnings

Output format

1. Script file at src/scripts/{name}.ts with --quiet accepted
2. Taskfile wiring with silent: true + {{.QUIET_FLAG}} (unless carve-out)
3. Test under tests/ if _lib/script_output is used
4. Linter output showing 0 FAIL

Gotchas

• Forgetting --quiet — the silent Taskfile layer wraps the script and the call fails with unrecognized arguments: --quiet
• Gating ❌ failures behind --quiet — failures must always print
• Using raw print() for progress lines — drops on minimal, no inheritance
• Adding silent: true to a release / install-keys task — bypasses the Hard Floor confirmation
• Editing Taskfile.yml's QUIET_FLAG var — single source of truth, do not duplicate
• Forgetting that import … from './_lib/script_output.js' resolves relative to the script's own directory — copy the import path from a peer under src/scripts/

Frugality Standards

Apply the Frugality Charter to every script you author. Phase 10 of the charter (settings hooks row — verbosity.script_output / verbosity.taskfile_command_echo) is what this skill exists to teach.

Examples in this artifact:

• Per the charter's default-terse rule, success lines are gated behind --quiet so the only thing visible at minimal is the end-of-run summary or a failure.
• Per the post-action summary suppression, scripts with multi-step output collect via success() and emit one flush_summary() line.
• Per the cheap-question check, scripts never prompt unless they hit a Hard Floor surface (§ 3 carve-outs).

Pre-save self-check:

1. Does every print("✅ ...") line sit behind --quiet / the helper?
2. Does the script add silent: true + {{.QUIET_FLAG}} to its Taskfile entry (unless a carve-out)?
3. Are failure lines (❌, exit non-zero) never gated by quiet?
4. Are Iron-Law surfaces using plain print() and not the helper?

Do NOT

• Do NOT gate ❌ / failure output behind --quiet
• Do NOT use _lib/script_output for release confirms or secret prompts
• Do NOT add silent: true to carve-out tasks
• Do NOT hardcode print() for progress in new scripts — use info()
• Do NOT skip the Taskfile wiring — without it the verbosity gates leak
• Do NOT edit dist/agent-src/, .augment/, or .claude/ projections

Cloud Behavior

On cloud surfaces (Claude.ai Web, Skills API) the package's task runner and _lib/script_output are not reachable. The skill still applies — with prose-only validation:

• Emit the full script + Taskfile snippet as copyable Markdown blocks. Do not attempt to write to disk.
• Self-check: --quiet accepted, failures never gated, success lines gated, helper imports look syntactically right.
• Tell the user to save under src/scripts/{name}.ts, wire the Taskfile entry, and run task lint-skills && task ci locally before committing.
• Skip every reference to running the linter or task commands yourself — they only run on the user's machine.

Examples

Good description (trigger-shaped, names domain + symptoms):

"Use when adding or editing any script under scripts/ — --quiet flag, _lib/script_output helpers, silent Taskfile wiring, Iron-Law carve-outs — even when you just say 'add a check script for X'."

Bad description (vague, no trigger):

"Script conventions"