BrandForge is a Codex skill for developing, testing, and preliminarily screening distinctive commercial brand names. It replaces unstructured brainstorming with a staged multi-agent process that separates invention from evaluation and keeps every recommendation tied to explicit evidence.
Use it to name or rename a company, product, app, software service, or platform when you need more than a long list of plausible words.
- Turns a product and audience brief into a sound and naming charter.
- Explores three deliberately different naming territories, each with a semantic source, construction technique, and phonetic profile.
- Uses independent agents for high-volume generation, linguistic review, semantic red-teaming, commercial research, trademark screening, domain research, and final review.
- Persists every explicitly proposed candidate, its journey, findings, and evaluation in a project-local Markdown registry before filtering or deduplication.
- Tests pronunciation, spelling, dictation, word of mouth, visual form, and risk in the priority launch languages.
- Generates approximately 100 raw candidates in coherent territories and phonetic subfamilies before narrowing the pool.
- Uses the Instant Domain Search MCP server for early bulk availability triage, then separately inspects occupied important domains for similar-project use on the finalists.
- Produces a traceable shortlist with evidence, confidence levels, unresolved risks, and next steps.
BrandForge is intentionally not a casual name generator. It does not create character names, lightweight title ideas, or pages of random pseudo-tech names.
The skill advances candidates through explicit stages:
- Brief — define the offer, audience, markets, languages, category, constraints, and desired personality.
- Sound and naming charter — establish what the name must and must not feel like, then define exactly three naming territories.
- High-volume independent generation — produce roughly 100 raw candidates across six isolated generators, three territories, and labeled phonetic subfamilies.
- Early domain triage — bulk-check exact standard-set domains, natural domain hacks, and contextual fallbacks with Instant MCP without biasing blind reviewers.
- Two-pass blind review — reduce the pool through high-volume triage and then a deeper linguistic and semantic gate.
- Focused iteration — evolve only the strongest subfamilies, check the new domains, and repeat the blind review.
- User gate — ask the user to mark each survivor as
advance,hold, orrejectbefore commercial research begins. - Commercial research — inspect current companies, products, apps, software, and close name variants for the user-approved candidates.
- Usage and visual tests — assess real sentences, radio and repeat-back behavior, case variations, glyph ambiguity, and simple monogram potential.
- Final trademark and domain screening — use official trademark registries, refresh Instant MCP results, and inspect occupied standard-set and proposed fallback domains.
- Final review — reconcile hard gates, scores, confidence, evidence, and unresolved questions into a provisional recommendation.
The workflow may return no winner. A weak or materially risky candidate is never promoted merely to complete a shortlist.
Codex discovers standalone skills from .agents/skills directories in a repository and from the user-level ~/.agents/skills directory.
Install BrandForge for your user account:
mkdir -p ~/.agents/skills
git clone https://github.com/danieleperilli/brand-forge.git ~/.agents/skills/brand-forgeFor repository-scoped use, place or symlink this directory at:
<repository>/.agents/skills/brand-forge
Codex normally detects skill changes automatically. Restart Codex if the skill does not appear.
Invoke the skill explicitly with $brand-forge, or describe a commercial naming task that matches its scope.
Use $brand-forge to name a B2B writing platform for independent authors.
The initial markets are Spain and the United States. The name should feel adult,
human, calm, and internationally pronounceable. Avoid obvious AI language and
test this sentence: “I am writing my next book with ___.”
The skill begins by collecting the smallest viable naming brief. Helpful inputs include:
- what is being named and the value it creates;
- the primary audience and desired response;
- launch countries, languages, and likely expansion markets;
- the product category, business model, and relevant goods or services;
- desired and forbidden traits, sounds, letters, words, or associations;
- preferred or excluded naming modes, such as English words, compounds, blends, invented names, or international/reference names;
- naming constraints such as length, exact-domain requirements, extra required TLDs, competitors, or brand architecture;
- one natural sentence in which a customer would say the name.
Unknown details can remain provisional. Target territories and goods or services are required before meaningful trademark screening can begin.
A complete run can produce:
- an agreed naming brief and a combined sound and naming charter;
- the naming territories and phonetic subfamilies explored, including the lessons from each mode;
- a durable candidate registry covering every proposal, lineage, pronunciation, construction, phrase tests, risks, research, gates, and decisions;
- a finalist comparison with hard-gate status, weighted score, and confidence;
- direct source links and timestamps for commercial, trademark, and domain findings;
- a dossier for each finalist;
- eliminated candidates with decisive reasons;
- a ranked provisional recommendation, or an explicit recommendation of none;
- a blind real-user validation script and professional legal next steps.
Scores are calculated only after every hard gate passes. Any review, stop, or untested gate keeps the total at N/A so incomplete evidence cannot create false confidence.
BrandForge declares the Instant Domain Search MCP server as a Streamable HTTP dependency in agents/openai.yaml, using:
https://api.instantdomainsearch.com/mcp/streamable-http
Immediately after high-volume generation, the skill checks every unique active candidate across .com, .io, .app, .net, .org, .chat, .to, .one, .co, and .ai, plus any TLD required by the brief. It also tests natural semantic or spelling domain hacks such as love.me or mim.ar; country-code TLDs are not added merely because they match a launch market. If the exact domains do not yield a credible path, it tests a small context-specific set of modified labels such as theprose or myprose.
This early result is provisional: it can stop a name only when an explicit exact-domain requirement fails, and it stays hidden from blind linguistic and semantic reviewers. New names produced during focused iteration receive the same check.
For every finalist, the skill refreshes all retained results and inspects the occupied exact-name standard-set domains, plus any domain hack or modified label in the proposed plan, to determine whether they are used by active similar or adjacent projects. The final domain gate can pass because Instant MCP currently reports a strategically credible domain as available without contradictory purchase-status evidence, or because the occupied exact-name standard-set domains have no confusingly similar active use and the brief does not require acquiring the exact domain.
If Instant MCP is unavailable, BrandForge leaves the domain gate untested and does not calculate a final score. It does not silently replace the required availability source with a generic web or WHOIS lookup.
BrandForge stores the registry in the repository of the product being named: <project-repository>/brand/brand-registry.md. It resolves the target project Git worktree before creating the file; it does not use the BrandForge skill checkout as an implicit fallback. The file opens with one human-readable table containing every proposed name, its last Stage, status, and update date, sorted by Stage descending so the most advanced candidates appear first. Each name is a link to its detail card below, where its motivation, pronunciation, findings, evaluation, stop reason or next action, and full journey remain easy to inspect. The coordinator writes every exact candidate immediately after it is proposed and before deduplication or evaluation.
The target repository's .gitignore receives precise rules for the registry, its lock, and temporary replacement files. This keeps the project-local record out of Git by default while leaving other brand/ assets available to the project. BrandForge never stages, commits, pushes, publishes, or syncs it without explicit authorization. It stores only the minimum project context needed to interpret an evaluation and must not contain credentials, signed/session URLs, personal data, or named test participants.
An older registry in the Codex data directory is not copied automatically. On explicit request, BrandForge can migrate one selected project record into the target repository while retaining the original source.
BrandForge performs preliminary research, not legal clearance. It never claims that a name is legally available, registrable, globally safe, guaranteed to be purchasable as a domain, or preferred by real users.
Trademark findings use one of three outcomes:
no evident conflict found in preliminary screening;legal review required;stop.
Comprehensive clearance, common-law rights research, likelihood-of-confusion analysis, filing strategy, and trademark applications require a qualified professional. Domain and marketplace observations are time-sensitive and should be rechecked immediately before action.
When BrandForge is used for a product repository, it creates this local, Git-ignored artifact:
<target-project>/
├── brand/
│ ├── brand-registry.md # Permanent candidate history; Git-ignored by default
│ ├── brand-registry.md.lock/ # Temporary cross-task lock
│ └── .brand-registry.md.*.tmp # Temporary atomic-write file
└── .gitignore # Precise rules for the three paths above
The structure below is the standalone BrandForge skill repository, not the target product repository:
.
├── SKILL.md # Workflow, stage gates, and agent roles
├── agents/
│ └── openai.yaml # Codex metadata and Instant MCP dependency
├── references/
│ ├── evaluation-framework.md # Hard gates, scoring, and confidence rules
│ ├── naming-modes.md # Semantic sources and construction techniques
│ ├── output-contract.md # Human-readable registry and deliverable formats
│ └── research-protocol.md # Commercial, domain, and trademark research
└── brand-forge.code-workspace # Optional VS Code workspace
The skill has no build step. Current domain availability checks require the Instant MCP dependency declared in agents/openai.yaml.
- Real multi-agent work, never simulated specialist roles.
- Strict separation between generation and evaluation.
- High-volume exploration before progressive blind filtering.
- Naming territories and phonetic families instead of undifferentiated name lists.
- Hard failures override numerical scores.
- User approval before commercial research.
- Current sources, direct URLs, timestamps, and explicit confidence.
- Instant MCP availability evidence kept separate from active website and market-use evidence.
- Durable, append-only candidate history with a separate current evaluation.
- Honest uncertainty and a preference for no recommendation over a weak one.
The complete operating procedure is defined in SKILL.md.