Unmachined

unmachined

Make it read written and look made, not generated.

AI output converges on the statistical center: the same vocabulary, the same cadence, the same purple gradient on the same centered hero. This skill holds two lines of defense. First, catalogs of known tells with severity tiers. Second, a variety requirement, because the fix becomes a new tell when every output converges on the same alternative. Never swap one monoculture for another.

Priorities: text first, design and layout equal second, branding last.

Always-on (native enforcement)

Once installed, the skill can stay on for the whole CLI session so agents do not write AI-text slop unless the user opts out.

Full contract: references/always-on.md.

Onboarding (required once)

On first use, or when onboarded_at is missing, ask:

Turn unmachined always-on for this CLI so all user-facing writing is anti-slop by default? Or leave off (only when you invoke the skill)?

Then save via:

python3 scripts/onboard.py                 # interactive
python3 scripts/onboard.py --always-on     # native on, text+ui
python3 scripts/onboard.py --always-on --text-only
python3 scripts/onboard.py --always-off
python3 scripts/onboard.py --show          # status

Config: project .unmachined.json overrides ~/.config/unmachined/config.json. Env UNMACHINED_ALWAYS_ON=1|0 overrides the boolean for a single session.

Agent duties when always_on is true

1. Resolve config (scripts/onboard.py --show or read the JSON files).
2. Load references/text-tells.md before drafting user-facing prose.
3. Write clean; do not draft slop intending to clean later.
4. Run scripts/scan_text.py on the deliverable; score >= threshold blocks ship.
5. If ui is in surfaces and the task is UI, also load design refs and scan_ui.py when editing UI source.
6. First content write of the session: one short line that always-on is active.

Out of scope: pure logic without user-facing strings; this skill's ban-list docs; raw logs and data dumps. See always-on.md.

Verbs

Hosts without slash commands use the same words in natural language.

Dispatch contract

1. Parse an explicit verb first. If none is present, use the default build workflow. A file named audit or fix remains a target when supplied as a path.
2. If verb is onboard / always-on / always-off / status, handle config only (shell scripts/onboard.py …). If verb is learn (or learn status|on|off), shell scripts/learn.py and load references/personal-voice.md.
3. If no config onboarded_at and the task will write user-facing text, run onboarding once before continuing (or instruct the user to run the script).
4. If always_on is true, apply the always-on agent duties even when the user did not type /unmachined.
5. Resolve the smallest target that satisfies the request. For diff, use the working-tree diff by default.
6. State detected surfaces in one short line when both text and UI apply.
7. audit and diff are read-only unless the user asks to fix.
8. Load only the references named by the chosen pipeline.

Load budget

One hop per level: description -> this file -> the reference the pipeline names. Bulk-loading references/ is a defect, not thoroughness.

• Text (default): text-tells.md only. Add voice-and-copy.md for microcopy, technical-founder-correspondence.md for founder or sales replies.
• UI: design-tells.md. Add design-system.md only when building or restyling, audit-playbook.md only for audit/fix on existing UI, and stack-rules.md only when package.json shows Next.js, Tailwind, or shadcn.
• Config verbs (onboard, always-on, always-off, status): scripts only, no references.
• Personal voice (learn verb, or fix/rewrite of the user's own prose when a profile is enabled): personal-voice.md only.
• Execute the scanners, never read their source. scan_text.py and scan_ui.py are run targets, not reading material.

Text pipeline (main priority)

1. Read references/text-tells.md. Microcopy also needs references/voice-and-copy.md.
2. Draft or edit under the rules (em-dash ban; no critical vocabulary; concrete claims; honesty gate).
3. Run the deterministic scanner before self-judgment: python3 scripts/scan_text.py <file> [file...] [--json] [--threshold 40] [--mode prose|chat|ui] Stdin works for pasted content. Score >= threshold blocks fix / build / always-on delivery. Fix criticals first. For JSX/TSX or other source files, do not treat a whole-file scan_text.py score as a copy verdict: operators, comments, template expressions, and placeholders can produce false positives. Run scan_ui.py on the source, then scan extracted user-facing strings or changed copy separately with scan_text.py. When a personal voice profile is enabled and the text is the user's own writing, add --personal and follow references/personal-voice.md for fix/rewrite exemplar targeting.
4. Self-critique once, naming what still reads as AI-generated. Revise for cadence and variety. Do not sand every sentence to the same length.
5. Honesty gate: every number is real or a labelled placeholder. No invented metrics, quotes, or anecdotes.
6. For technical founder DMs, prospect chats, or sales-engineering replies, load references/technical-founder-correspondence.md. Separate truth approval from voice editing: the domain owner confirms the substance first, then improve the wording without changing claims. For outreach batches, scan every message individually and scan the concatenated bodies as a batch; repeated openers, questions, founder lines, CTAs, or sentence rhythms block shipment even when each message passes alone. Deleted or retracted messages remain history, but their claims and CTAs are inactive until a corrected replacement is sent.

Design pipeline

1. Read references/design-tells.md and references/design-system.md. Audits also need references/audit-playbook.md. python3 scripts/scan_ui.py <src-dir> [--json] [--threshold 40] [--tailwind-version auto|3|4]
2. For Next.js/Tailwind/shadcn, read references/stack-rules.md and match the project pin.
3. Structure before color; one justified aesthetic risk; WCAG 2.2 AA; reduced motion; focus visible; 320-768px layouts.

Cross-output variety

Consecutive outputs must not share the same layout skeleton, display font, or accent hue unless the project design system requires it.

Resolved rule conflicts

• Em-dash: banned in generated text. Hyphen, comma, or restructure.
• Serif type: only with justification; not a lazy default.
• Metrics: real, cited, or labelled placeholder.
• Eyebrows / section numbers: only when they encode a true sequence.

Reference index

Output discipline

This skill's own reports must pass its own rules. Findings are concrete: quote the line, name the rule, show the fix. No praise padding. Never claim a scan, render, or test that did not run.