Canonical voice
Voice.

Paste the markdown into an AI tool before drafting Evalgist copy. It is generated from @evalgist/content-style.
# Evalgist content style guide

Copy this entire guide into your AI agent as context before writing any Evalgist copy — landing pages, guides, blog posts, product text, emails, or sales material. It is the current, authoritative voice. Do not invent rules beyond it.

This guide is generated from `@evalgist/content-style`. Do not edit it by hand; edit the constants and regenerate.

## What Evalgist is

Evalgist builds AI tools that help professionals evaluate a batch of structured documents against one criteria set — faster, more consistently, and with evidence. The human remains the decider; the AI supplies substantiation, not conclusions.

Three-beat distillation: One criteria set · Quoted evidence · Exportable record.

The human is always the decider. The AI supplies substantiation, extraction, comparison, sorting, summary, or citation — never the conclusion.

## The job being sold

Evaluate a batch of documents against one criteria set, and keep an evidence trail.

It replaces: Ad-hoc chat-assistant triage: one document at a time, criteria that drift between prompts, and no record when the decision is questioned.

What the user gets:

- one criteria set applied to every document
- a quote from the source behind every judgment
- an exportable record for the decision meeting

Sell: Speed, consistency, and an evidence trail — not a smarter model.

## Who it is for

Hiring managers and selection panels evaluating a batch of applications against the same criteria.

Segments: selection committees, hiring managers at SMBs and non-profits, academic and public-sector panels, grant and admissions reviewers.

Academic hiring is where the workflow was first built and proven. Treat it as origin and channel, never as the only intended user in marketing copy.

Not built for: teams that need deep ATS integration; fully automated decision pipelines.

## Contrasting with ad-hoc chat workflows

Contrast with the ad-hoc chat workflow, not with a named vendor or model. The wedge is workflow, not intelligence.

The three contrasts:

- Criteria drift: a chat re-interprets the standard with every prompt; a criteria set holds it across the whole batch.
- No quotes: a chat summarises; an evaluation cites the source line behind each judgment.
- No packet: a chat leaves a scrollback; an evaluation leaves a record you can bring to the meeting.

Approved patterns:

- `A chat can score one resume. It will not hold one standard across two hundred, or hand you a packet for the meeting.`
- `A chat window starts every conversation from zero. A criteria set holds the same standard for every document in the batch.`
- `You can ask a chat why it ranked someone. You cannot hand its answer to the committee.`

Forbidden patterns:

- `Better than ChatGPT.`
- `Smarter than a chatbot.`
- `Don't trust generic AI.`

## Selling vs telling (marketing surfaces)

Sell the job, not the disclaimer. The reader is drowning in documents and has a review to finish; meet that first, and let the trust surface carry the caution.

- Lead with the pain and the artifact — the batch, the quote, the record. Never with controls, storage, retention, or provider names.
- Compliance is furniture, not a feature. EU storage, ZDR routing, and retention windows belong on trust pages and in the FAQ; on a marketing page they get at most one line with links.
- Never invent segment jargon that needs a glossary. Describe the situation instead: hundreds of applications, one criteria set.
- Do not pre-reject buyers in the funnel. Scope boundaries ('not built for …') belong in the FAQ, not in heroes or feature copy.
- The origin story is one line on marketing pages and a section on the about page. It never becomes the target audience.
- Say 'the human decides' once per page. Repeating it reads as apologizing for the product.
- Use the record as the payoff, not the mood. Name the evidence trail first. Say 'defend' or 'defensible record' only when the full source survives the relevant retention window.
- Every section must earn its place against the 10-second test: what job does this do, what do I get out, can I try it now. Cut or demote everything else.
- Footers are navigation, not a legal register. Account-level trust links only; product legal lives with the product surface.

Smell test: If a section would fit a procurement portal or a privacy notice better than a product page, it is in the wrong place.

## Defend and defensible claims

Originals and extracted text are kept only for a 14-day recovery window. Evidence quotes and exportable summaries remain until the customer deletes the job or account.

Say defend, defensible, or defensible record only when the full source document survives the relevant retention window.

Allowed for the current recovery-window posture:

- evidence trail
- audit trail
- anchored to source at evaluation time
- exportable record

Do not use without source retention:

- defend
- defensible
- defensible record

## Tone

Sober, confident, slightly dry. Short sentences. The Economist, not TechCrunch. English as the primary language, even in a Belgian context.

Never use:

- emojis
- exclamation marks
- superlatives
- startup enthusiasm ("we're so excited", "the future of…")
- irony or provocation toward regulators

## Product naming

Pattern: `Evalgist <Product>`. Products are descriptive suffixes under the Evalgist parent, never sub-brands with independent identity. The set: Shortlist, Grade, Extract.

- First mention on any external surface uses the full name: `Evalgist Shortlist`.
- Second mention may shorten to the suffix: `Shortlist`.
- Inside the signed-in product, the sidebar lockup is the first mention; body copy may then use the suffix as a proper noun ("Shortlist is processing…") or as a common noun for the deliverable ("Start your first shortlist").

When in doubt: outside the auth boundary, full name on first mention; inside it, the lockup carries first mention.

## Words

Prefer:

- evidence
- quoted evidence
- one criteria set
- batch
- evidence trail
- audit trail
- exportable record
- evaluate
- shortlist
- grade
- consistent
- fair
- faster
- per criterion
- per candidate
- per answer
- review
- decide
- verify
- professional
- human decision

Avoid:

- revolutionary
- game-changing
- unleash
- supercharge
- seamlessly
- effortlessly
- 10x
- cutting-edge
- next-generation
- disruptive
- bias-free
- objective AI
- fully compliant
- AI sparkle
- magical

## How to talk about AI

Name the function, not the marketing claim. AI is the means, not the product.

- Good: `AI-assisted evaluation`
- Good for SEO: `AI resume screening with quoted evidence`
- Good for in-product status: `AI-assisted, sorted by match`
- Bad: `Powered by AI`

### Multi-pass analysis tiers

An extra pass tells the reviewer where to look. It does not make the suggestion more correct, and it never decides.

- Never claim independence between passes. Passes run on related model families and share blind spots; describe them as cross-checked, never as independent reviewers or as models that 'agree'.
- Never imply an extra pass improves the score, grade, or ranking. Sell what it produces: the criteria worth the reviewer's time.
- Name the tier after the process, not the outcome. 'Cross-checked' describes what ran; 'Verified' asserts a result the product cannot stand behind.
- Never name a model in public copy. A model is a config value that changes; the processor behind it is named in the subprocessor register, which is where that question belongs.

Approved:

- `A second pass challenges every judgment and flags the ones that do not hold up.`
- `Cross-checked analysis tells you which twelve of four hundred criteria to look at.`
- `The extra pass does not change the suggested mark. It changes where you spend your attention.`

Never:

- `Two independent AI graders agreed.`
- `Verified by a second model.`
- `Higher accuracy with multi-model consensus.`
- `Powered by <model name>.`

### AI search terms

Evalgist competes for real search intent around AI-assisted evaluation. Using the word "AI" is fine; AI theatre is not.

Allowed terms — use in title, description, summary, H1, H2, early body copy when the page actually answers that intent:

- `AI resume screening`
- `AI-assisted resume screening`
- `AI candidate screening`
- `AI-assisted evaluation`
- `AI document review`
- `AI-assisted document review`
- `AI grading`
- `AI-assisted grading`

Binding rule: When an AI search term appears, bind it in the same sentence or the next sentence to a concrete function: extract, compare, sort, summarize, cite, or produce an evidence trail.

Good patterns:

- `AI resume screening with quoted evidence. Every match cites the resume line it rests on.`
- `AI resume screening with quoted evidence.`
- `AI-assisted evaluation that keeps the human as the decider.`
- `AI document review that cites the source line behind each judgment.`
- `The AI extracts, compares, and cites evidence. The reviewer decides.`

Bad patterns:

- `AI-powered hiring that transforms screening.`
- `Powered by AI.`
- `Objective AI candidate screening.`
- `Fully compliant AI evaluation.`
- `Automated hiring decisions.`

Rule of thumb: Use AI terms for discoverability. Do not use AI as the value proposition.

## Rules for every surface

Always:

- Anchor evaluation language in evidence, criteria, and human decision-making.
- Label AI suggestions and human decisions as different states.
- Name AI as the method only when it explains a visible function.
- Use SEO-visible AI terms when they match real search intent, then bind them immediately to criteria, quoted evidence, records, or human review.
- Use the full product name on first mention outside the authenticated product frame.
- Treat every trust claim as a public claim that may be read by a customer, candidate, auditor, or regulator.

Never:

- Promise bias elimination, objectivity, complete compliance, or automated correctness.
- Use startup acceleration language to make evaluation work sound effortless.
- Make the customer, candidate, regulator, or protected group the butt of the joke.
- Replace a concrete source, quote, criterion, or workflow detail with generic AI capability language.

### Claim boundaries

- Bias: Say reduce avoidable bias, improve consistency, or make evidence visible. Do not say bias-free.
- Compliance: Say evidence trail, audit trail, or GDPR/EU AI Act caution for the current recovery-window posture. Say defensible record only when the full source survives the relevant retention window. Do not imply certification or full compliance from copy alone.
- Speed: Tie speed to a workflow artifact: faster review, faster comparison, faster evidence lookup. Do not imply the human judgment disappears.
- AI: Tie AI to a function: extract, compare, sort, summarize, cite. Do not use AI as a standalone value proposition.
- SEO: Put AI search terms in titles, descriptions, headings, and early body copy when relevant. The next clause or sentence must explain what the AI does and what the human still decides.

## Compliance posture

External copy avoids irony or provocation toward regulators. Reason: GDPR + EU AI Act audit-caution. No wordplay around pretending, faking, cheating, or fraud. No absolute fairness or full-compliance claims.

## Citing claims

The stricter the claim, the more explicit the backing.

- Cite or link: statistics, study findings, legal claims, regulatory expectations, scientific claims, market facts, named authorities.
- Soften or remove: claims you cannot support quickly — especially "bias-free", "objective", "fully compliant", "guaranteed", "automatic".
- Attribute precisely: "Ladders reported 7.4 seconds in its 2018 eye-tracking study" beats "research shows recruiters spend seconds".
- Keep product claims bounded to what the workflow produces: evidence quotes, criteria comparisons, records, exports, retained or deleted data, human-review controls.

## Landing pages

### Hero formula

- Eyebrow: full product name + a trust beat, 50 characters max
- H1: 7 words max, declarative
- Subhead: 14 words max, evidence-led
- CTA: 4 words max, verb + object, no period
- Fineprint: 2-3 trust beats joined by `·`

### Closing CTA formula

- Eyebrow: "Built for" or equivalent
- Body: 20 words max, evidence-led, ends with the audience constraint
- CTA: 4 words max

### Trust beats

Approved fineprint, shown next to a CTA (each is a public claim):

- No payment details required
- No subscription
- GDPR-compliant
- EU data residency
- Every result backed by evidence

## Long-form: pick a register first

### Guides — evidence-first practical guide

Use for: guides, how-to pages, evergreen explainers. Do not use for: dated blog essays, launch notes, legal pages, in-product task copy.

Evergreen, practical, search-oriented. The reader should leave with a method they can use without buying Evalgist.

- Opening: Open on the practical failure the reader recognises, then move quickly to the method. Example: "Most screening goes wrong before anyone reads a word."
- Evidence: Cite external research or authority when making broad causal, legal, statistical, or scientific claims. Plain workflow advice may stand without a citation.
- Structure: Use direct instructional headings. Each section should name an action, a trap, or a record the reader can produce.
- Product: Mention the relevant Evalgist product only after the useful method has been explained, and keep the CTA secondary.
- Ending: End on the professional standard the method creates, not on a sales flourish.
- Cadence: short sentences, paragraphs of one to three sentences, headings instructional, not editorial.

### Blog essays — contrarian-evidence essay

Use for: blog essays, LinkedIn-native posts, opinion-led newsletter notes. Do not use for: legal pages, regulator-facing copy, in-product task copy, evergreen guides.

Opinion-led and built to travel, but still sober Evalgist voice. Be provocative about weak evaluation practice, never about the law, the customer, or the person being evaluated.

- Opening: Open on a claim, not a definition. Example: "Your shortlist was mostly decided before the meeting. The rest was paperwork."
- Antagonist: Name the bad practice, never a protected group, regulator, customer, or individual.
- Evidence: Include at least one concrete study, number, or cited authority when making the central claim.
- Pull-quotes: Write two or three short, standalone sentences that can be quoted without context.
- Ending: End on a line the reader can repost as a statement of professional standards.
- Allowed edge: dry criticism of lazy evaluation; contrarian framing of common hiring habits; plainly naming bias hiding behind taste, fit, prestige, or speed.
- Forbidden edge: mocking regulators; claiming compliance as a joke; fraud or pretending wordplay; startup hype punctuation; humour that makes the customer look foolish.
- Cadence: short sentences, paragraphs of one to three sentences, headings argumentative, not instructional.

## Check before you ship

- Does the copy preserve the human-as-decider / AI-as-substantiation distinction?
- Are AI suggestions and human decisions labelled as different states?
- Does the copy name the evidence trail, and use defend/defensible only when the full source survives the relevant retention window?
- Is every broad factual, legal, statistical, or research claim either sourced or softened?
- If an AI search term appears, is it immediately bound to criteria, quoted evidence, records, or human review?
- Does the first external product mention use the full Evalgist <Product> name?
- Does any line imply bias-free, objective, fully compliant, or automatic decision-making?
- For a multi-pass tier: does the copy avoid claiming independence between passes, a better result, or a named model?
- Could the same sentence appear in a sober professional publication without exclamation marks, hype, or forced warmth?
- For guides: can the reader use the method without buying Evalgist?
- For blog posts: is the antagonist a weak evaluation practice, not a person, customer, regulator, or protected group?