laconic-skill

AI agents write too much.

laconic-skill checks AI output before humans have to read it. It can pass, fail, rewrite, and emit receipts for concise-output checks without calling a model during verification.

The model drafts. The runtime verifies.

laconic-skill is a deterministic verifier for concise AI output. It runs as a local CLI, emits receipts, and validates as a Claude Code plugin.

Use it for:

• AI-generated pull request descriptions
• code review comments
• agent status updates
• documentation drafts
• any workflow where verbose AI output wastes reviewer time

Why this exists

Prompting can ask an AI to be brief. It cannot prove the output stayed brief.

laconic-skill runs outside the model. It applies deterministic rules, returns pass/fail results, and can emit a JSON receipt for the check.

Try it in 60 seconds

Install and build once:

npm install
npm run build

Check a concise output:

node dist/cli.js check fixtures/pass/compliant.txt --receipt

Rewrite a verbose output:

node dist/cli.js rewrite fixtures/fail/verbose_recap_heavy.txt --receipt

Run the full pipeline:

node dist/cli.js pipeline fixtures/fail/filler_heavy.txt --task writing --receipt

Emit a receipt from stdin:

cat output.txt | node dist/cli.js check - --receipt

Verified proof

• 50 corpus files: the benchmark runs against realistic verbose AI outputs in benchmarks/corpus/*.txt.
• 63.477% average character reduction: the benchmark reports average shrinkage after deterministic rewrite.
• 50/50 fixable outputs passed after rewrite: every benchmark output that started failing passed after rewrite.
• 0 compliant false fails: short compliant controls in benchmarks/compliant/*.txt stayed passing.
• Deterministic across 5 runs: the benchmark repeats and compares stable output signatures.
• No model calls during verification: the verifier runs locally and deterministically; package.json includes no model SDK dependency.
• Receipt hash determinism: tests verify the same input/output/config produces the same receipt hash.
• Claude Code plugin validation: the plugin manifest passed claude plugin validate ..

What it does

A local verification runtime for concise AI output.

• checks output
• rewrites verbose output
• runs pipelines
• emits receipts
• optionally uses local style memory
• integrates with Claude Code

Architecture

input
-> draft text
-> laconic verifier
-> optional deterministic rewrite
-> final text
-> receipt

Core pieces:

• skills/laconic-responses/SKILL.md defines the laconic response behavior.
• src/verifier.ts checks length, bullets, filler, preambles, repeated prompts, answer-first opening, and caveat limits.
• src/rewrite.ts trims output without inventing facts.
• src/receipt.ts creates hash-bound receipts.
• src/pipeline.ts connects draft text, verification, optional rewrite, placeholder correctness checks, and receipts.
• src/memory/ contains optional local style memory.

Local memory boundary

Memory can guide style. Memory does not decide verifier pass/fail.

Memory can affect style retrieval, preferred phrasing, and rewrite suggestions. It does not override deterministic checks or rule enforcement.

Memory-enabled pipeline receipts may include memory metrics so the receipt identifies the exact run context. Core verifier receipts do not require memory.

Examples

• Full walkthroughs: EXAMPLES.md
• Basic check: examples/basic-check.md
• Rewrite before/after: examples/rewrite-before-after.md
• Pipeline receipt: examples/pipeline-receipt.json

Claude Code plugin

Local test:

claude --plugin-dir .

Invoke the skill:

/laconic-skill:laconic-responses

Validate the plugin:

claude plugin validate .

CLI usage

Check output:

node dist/cli.js check fixtures/pass/compliant.txt --receipt

Rewrite verbose output:

node dist/cli.js rewrite fixtures/fail/verbose_recap_heavy.txt --receipt

Run the pipeline:

node dist/cli.js pipeline fixtures/fail/filler_heavy.txt --task writing --receipt

Use local style memory:

node dist/cli.js memory add fixtures/pass/compliant.txt --outcome accepted --task writing
node dist/cli.js memory search "npm run build" --limit 5
node dist/cli.js pipeline fixtures/fail/filler_heavy.txt --task writing --memory --receipt

License

laconic-skill is open source under the Apache License 2.0.

Attribution is preserved through the project NOTICE file:

Electric Wolfe Marshmallow Hypertext | Tionne Smith, 2026.

The software is Apache-2.0 licensed. Project names, marks, and branding are reserved separately.

Inspired by Karpathy's public skill-file. Not affiliated with Andrej Karpathy.