nodecode-on-brand

On-Brand: brand voice and compliance rules a deterministic linter enforces. A Nodecode rewrite of Mhiah/On-Brand by Mhiah.

A prompt can ask a model to write in a brand's voice; it still slips — a "we" in a third-person brand, a banned buzzword, "guaranteed returns" in a crypto post, a missing NFA. On-Brand moves those rules out of the prompt and into the harness. Every draft is scored by a rule-based linter, so the same text always gets the same verdict and the model never grades its own work. While a brand is active its rules ride every model request, and a tool call that sends, posts or publishes off-brand copy is refused before it goes out.

Get it

Tell the agent "add the on-brand cell", or run nodecode add nc://on-brand.

Set it up

One required setting, the folder the profiles are kept in:

{
  "on-brand": {
    "enabled": true,
    "store_dir": "~/.nodecode/on-brand/",   // profiles and the check log, one file
    "gate": true,                           // refuse off-brand outbound tool calls
    "gate_patterns": []                     // tool names that send; the built-in list when empty
  }
}

store_dir has no default on purpose: the profiles are yours, kept where you say. Without it the cell refuses at start and names the member. The section is in /settings too. Nothing here reaches the network and there is no key.

Member	Default	What it does
`store_dir`	none (required)	the folder the brand profiles and the check log live in
`gate`	`true`	refuse a tool call that sends, posts or publishes copy breaking a hard rule
`gate_patterns`	the built-in list	the tool-name patterns that count as outbound (`send`, `post`, `tweet`, `publish`, `reply`, ...)

Use it

/on-brand starter        # two example brands to try
/on-brand use nimbus     # turn enforcement on for one
/on-brand check We are going to the moon!!! GUARANTEED returns. LFG

nimbus: 0/100: first person (we, we're) in a third-person brand; banned word
'to the moon'; banned word 'LFG'; 3 exclamation marks, limit is 0; Crypto /
Web3 promotion: promises guaranteed returns; Crypto / Web3 promotion: add 'NFA'
or 'DYOR' when talking about price, tokens or returns

Or just describe the brand: "Set up an on-brand profile for Acme Labs: third person, calm and precise, never say synergy or game-changer, max 280 characters, crypto compliance." The model calls onbrand_rules and saves it.

The four tools

Tool	What the model uses it for
`onbrand_check`	score copy 0-100 against a brand and list every rule it breaks; `fix=true` also answers a corrected version with the pinned spellings applied
`onbrand_rules`	create or update a profile: person, tone, banned words, preferred spellings, required phrases, packs, limits, examples
`onbrand_learn`	draft a profile from 2 or more approved posts: the person, the emoji, hashtag and exclamation ceilings, the exact spelling of the name and whether the crypto pack applies
`onbrand_voiceover`	turn on-brand copy into a script that reads aloud: no markdown or links, numbers and money spelled out, optional SSML pauses

The rules

A profile carries the person (third, first, any), a tone, banned words, exact spellings, required phrases, the character / hashtag / emoji / exclamation limits, sample lines, and compliance packs:

crypto: no guaranteed returns, no "risk-free", no "100x" or "to the moon", no urgent buy; NFA or DYOR once the copy talks price, tokens or yield.
health: no cure claims, no rapid weight-loss promises, no guaranteed results; "not medical advice" once the copy makes a health claim.

Hard rules (person, banned words, required phrases, length, compliance) fail a draft. Soft rules (spellings, hashtag, emoji and exclamation limits) lower the score and are repaired when the copy is fixed.

`/on-brand`

/on-brand                      status: active brand, the gate, the totals
/on-brand list                 every brand profile
/on-brand use BRAND            turn enforcement on for a brand
/on-brand off                  turn enforcement off
/on-brand show [BRAND]        a brand's rules
/on-brand new BRAND            create an empty profile
/on-brand starter              add two example profiles to try
/on-brand set BRAND FIELD VALUE  person, tone, display_name, max_chars,
                               max_hashtags, max_emojis, max_exclamations
/on-brand ban BRAND WORD ...   /on-brand unban BRAND WORD
/on-brand spell BRAND WRONG RIGHT
/on-brand require BRAND PHRASE
/on-brand pack BRAND crypto|health [off]
/on-brand check TEXT           score TEXT against the active brand
/on-brand fix TEXT             the pinned spellings, corrected
/on-brand log                  the last checks, fixes and blocks
/on-brand gate on|off          where the gate switch lives
/on-brand delete BRAND

The same verbs are callable through eval: (on-brand:check "text"), (on-brand:rules "acme"), (on-brand:learn "acme" '("...")), (on-brand:voiceover "text"); each answers a string, and each refusal is ERROR: ON-BRAND-ERROR with the reason.

What is enforced, and where

Every request. While a brand is active, a standing harness section carries its rules and asks for copy in ```copy blocks, so a first draft starts on brand instead of being corrected afterwards.
Every outbound call. A tool whose name sends, posts, publishes, replies or messages is checked at the :tool point: copy that breaks a hard rule fails the call, with every rule named, and the model rewrites and retries. Clean copy goes through; onbrand_* tools and reads are never gated.
The log. /on-brand log reads the last checks, fixes and blocks, with the score each scored.

Tests

.github/test-cell.sh tools/nodecode-on-brand

Fourteen tests, offline: the lifecycle and its refusals, the linter through the tool seam, a verdict that does not move, a profile that outlives a restart, /on-brand, the standing section, the gate, learn and the voiceover script. Nothing here dials anything: a check is a linter, not a service.