AI builds whatever you ask. Ask well.
No brief yet? Build it on the Directory, guided →
0 to 100, a weighted average over fourteen public rules. Problem, users, outcome, scope, budget, deadline.
Four hard requirements. Fail one and the brief is blocked, whatever the score. A brilliant brief that names the client stays off a blind noticeboard.
Every verdict quotes the brief. The model only judges; code owns every number, so the same brief and the same version always agree.
Two rules, anonymisation and the €10k floor, are noticeboard policy, not virtues. They carry context: directory; send gate_contexts: [] to score a generic brief without them.
Software is cheap to build now and easy to build wrong. The brief is the one part you can still get right before anyone, human or model, starts. 20% brief, 80% team: fourteen open rules check the 20%.
High quality: problem, users, metrics, definition of done, budget, timeline — all there.
The same brief names a payment provider and an email address. Quality is not admissibility.
One seam. The rules are versioned YAML files anyone can read and anyone can change. A model reads one rule at a time and answers pass, partial or fail, with a quote. Past the seam, code does the arithmetic.
The bar is fourteen files. Yours to change.
Every rule is a YAML file with a weight, a criterion and a source. Propose one, reweight one, argue one away — a pull request is the whole process, and it merges on evidence rather than on our say-so. What survives here is what we run on welance/Directory.
It started with briefs for digital products. Whether the same bar should hold for other kinds of work is precisely the sort of thing to open a rule change about.
The judge runs at temperature zero, but model output is not promised to be deterministic across provider updates. Service-funded verdicts are cached by ruleset version, model and brief; bring-your-own-key calls are always fresh. A score exists only when the judge returns all fourteen valid verdicts, each with evidence copied verbatim from the brief. A corpus of labelled examples runs in CI: a rule change that moves the numbers has to move those fixtures too.
welance runs authoritative publish scoring on DeepSeek V4 Pro. Suggestions and their verifier run on DeepSeek V4 Flash. The model names are returned by the API; Sonnet is not enabled on welance deployments.
curl -X POST https://briefs.welance.com/v1/score \
-H 'content-type: application/json' \
-d '{"brief":"…","judge":"llm"}'
const res = await fetch("https://briefs.welance.com/v1/score", {
method: "POST",
headers: { "content-type": "application/json" },
body: JSON.stringify({ brief, judge: "llm" }),
});
const { score, decision, gate } = await res.json();
import httpx
r = httpx.post("https://briefs.welance.com/v1/score",
json={"brief": brief, "judge": "llm"})
score, decision = r.json()["score"], r.json()["decision"]
{
"score": 92.0,
"band": "Directory-ready",
"decision": "blocked",
"gate": {
"passed": false,
"missing": ["anonymised"]
},
"verdicts": [ … one per rule: status, confidence, verbatim evidence … ],
"review_required": false,
"ruleset_version": "1.0.0+83107baec655"
}
ruleset_version pins which bar judged a brief; the quotes keep it readable after the model is gone. The console works offline too, on a keyword stub.
{"mcpServers": {"brief-bar": {
"command": "uvx",
"args": ["--from", "git+https://github.com/welance/perfect-brief#subdirectory=mcp-server", "brief-bar-mcp"]
}}}
curl -X POST https://briefs.welance.com/v1/score \
-H 'content-type: application/json' \
-d '{"brief":"Paste your brief here"}'
briefs.welance.com/docs
briefs.welance.com/console.html
git clone https://github.com/welance/perfect-brief.git
All four work without an API key. Every endpoint, the MCP tools and the ruleset format are on the integrate page.
We are welance, a collective of independent senior teams since 2011. We run welance/Directory, a blind noticeboard where teams pick up briefs on the strength of the brief, not the brand. Its guided builder asks these fourteen questions and scores every draft with this service.
One company running both the noticeboard and the bar is a conflict. Our answer is to make the bar open: public rules, public weights, a public test corpus, MIT. Disagree? Fork it.
welance also pitches as a team on the Directory. Same fees, same blindness, same bar; the version on every verdict is how anyone checks. What binds us is the Operator Covenant: never mining briefs for leads, never training on them.
This tool is the distilled version of what fifteen years of handoffs taught us: the products that go well aren't the ones with the most polished decks — they're the ones where a small, senior, cross-functional team got a brief honest enough to start from. We can't open-source the teams. We can open-source the bar.
Together, this service and the Directory are the digitalised version of what welance has always done in person — matching honest briefs with senior teams — rebuilt with the approaches we now believe in: open rules instead of opinions, inspectable scores instead of gut feel. Two convictions drive it. Writing a good technical brief for a software product should be easy — the bar should help you clear it, not just measure you. And no one should have to commit to the first team they meet: a brief worth building deserves to meet several, so the fit — technical and cultural — is chosen, not defaulted to.
Details in GOVERNANCE.md and CONTRIBUTING.md. The long-form reasoning is in the article.