· Lean Edition Enterprise edition →
06 · Guardrails

Guardrails, not gates.

Enterprises slow content down with approval committees, RACI charts, and three sign-offs per post. You don't have the people for that, and you don't need them. Replace the committee with a few automated, always-on guardrails that run on every draft, catch the things that actually hurt you, and let everything else ship.

Owner
1 person + agents
Sign-offs
0 standing committees
Checks
Automated, pre-publish
Speed
Seconds, not weeks
Mindset shift

Gates → guardrails

A gate is a person who has to say yes before you move. A guardrail is a rule the system enforces while you move. Gates mostly protect against blame. Guardrails protect against actual harm, so pick those.

Enterprise gates
Legal, brand & comms each review every asset before publish
RACI chart with 6 roles per content type
Weekly governance meeting to unblock the queue
Approval lives in someone's inbox & memory
Throughput capped by the slowest reviewer
Lean guardrails
One brand-as-prompt spec the AI always runs under
Automated eval suite runs on every draft in CI
A pre-publish checklist the owner clears in a minute
Rules live in version control, visible to everyone
Throughput capped only by your ideas
Field note

Keep exactly one human gate: the owner's final glance before publish. Everything upstream of it is automated. That single gate is cheap because the guardrails have already done the heavy lifting.

Guardrail 1

Brand-as-prompt

Stop treating your style guide as a 40-page PDF nobody reads. Make it a system prompt the model loads on every generation. Encode voice, editorial rules, banned phrases, and a glossary once. Version it in git. Every agent inherits it automatically.

# brand.spec.md, loaded as system prompt by every content agent VOICE "Concrete, opinionated, no hedging. Short sentences. Active voice." Read as a sharp founder, never a brochure. ALWAYS - Lead with the claim, then the evidence. - Reading level: grade 8 to 10 (Flesch-Kincaid). - US spelling. Oxford comma. Sentence-case headings. NEVER - Banned: "revolutionary", "game-changer", "synergy", "leverage" (verb) - No unverified stats. Every number must cite a source in /facts. - No first-person plural fluff ("we believe", "we're excited to"). GLOSSARY guardrail -> automated, always-on check (NOT "gate" / "approval") agent -> an AI worker with a defined task + tools the product -> always "Atlas", never "the platform" / "our solution" DISCLOSURE If AI drafted >50% of the asset, add tag: ai-assisted: true
VERSIONED IN GIT ONE SOURCE OF TRUTH INHERITED BY ALL AGENTS DIFF-REVIEWABLE
Watch-out

A prompt drifts the moment your brand does. Treat brand.spec.md like code: PR to change it, one owner approves, changelog kept. Don't let three people quietly fork their own voice.

Guardrail 2

Automated evals before publish

Before any draft ships, an eval suite runs against it, the same way tests run against code. It checks that the draft is on-brand, grounded in your sources, free of hallucinations, and at the right reading level. A runner like promptfoo (illustrative, not an endorsement) can wire these into CI so a failing check blocks merge. Below is the owner's live pre-publish checklist. Toggle each item.

Pre-publish checklist

0of 5 checks passed
2 of 5 checks must pass before publish is unblocked.
Guardrail 3

Disclosure & provenance, kept light

You don't need a compliance department. You need honesty where it matters and a record you can reconstruct. Label AI-assisted content where a reader would care, and keep a one-line source trail per asset. That's it.

Disclose where it matters

  • Do label AI-assisted thought-leadership, customer-facing claims, and anything attributed to a named person.
  • Skip the badge on routine product copy or internal drafts. Over-disclosure just adds noise.
  • Use a single, plain tag: "Drafted with AI, reviewed by [name]."

Keep a simple source record

provenance.log ──────────────── asset: blog/atlas-launch model: claude-opus spec: brand.spec.md@a91f2 sources: /facts/pricing.md /facts/case-acme.md ai_pct: 72% reviewer: luke evals: 5/5 pass
Why this is enough

A spec hash, a source list, and a reviewer name let you answer "where did this claim come from?" in seconds. That's the only governance question that actually matters when something gets challenged.

Watch-list

The five risks that matter

Ignore the long enterprise risk register. For a lean content engine, these are the failure modes that actually bite, each paired with the one guardrail that contains it. Amber = manage, red = block on it.

Ungrounded claims / hallucination
Mitigation: RAG grounding + hallucination eval block merge; every number cites /facts.
Likelihood × Impact
Off-brand at scale
Mitigation: brand-as-prompt + on-brand eval on every draft; voice can't drift silently.
Likelihood × Impact
Stale / inaccurate source content
Mitigation: single structured source of truth; flag /facts older than a set TTL.
Likelihood × Impact
Over- or under-disclosure
Mitigation: one disclosure rule in the spec + provenance eval; consistent, not case-by-case.
Likelihood × Impact
Prompt / data leakage
Mitigation: no secrets in prompts; scoped API keys, redaction on inputs, vendor data-retention off.
Likelihood × Impact
Right-sizing

Five risks, five guardrails. If a control doesn't map to one of these, you're adding weight that slows a lean team down. Add controls only when a risk graduates from amber to red.

06 · done

Before you move on

Guardrails are the thing that lets a one-person engine ship fast without shipping something off-brand or untrue. Make sure these are in place before you scale volume, everything downstream assumes a draft can't reach the public unless it has already cleared them.

What you should have now

  • Brand rules as machine-readable context, voice, banned phrases and glossary in a versioned brand.spec.md that every agent loads.
  • An automated eval suite that gates publish on accuracy and on-brand voice, grounding, hallucination, readability and disclosure checks that block merge when they fail.
  • Risk tiers you act on, the few amber/red failure modes named, each mapped to the one guardrail that contains it, with nothing extra slowing you down.
  • One human sign-off where it matters, the owner's final glance plus a light provenance record, so you can always reconstruct where a claim came from.

Why this sets up the next stage

Guardrails are what make it safe to scale: once checks run on every draft automatically, you can turn up volume without turning up risk, because nothing reaches the public until it has passed.

With safety handled, the open question becomes output. Next, Done & Measure defines what “done” actually looks like and picks the few metrics that prove the whole system is working.

Next chapter
07 · Done & Measure
What "done" means, and the few metrics that prove the system works.