Context is the difference between AI that writes and AI that writes like you.
Sempleo treats context as infrastructure — curated, versioned, reviewed, and attached automatically to every agent run. Five layers. One coherent system. No prompt gymnastics.
A prompt is the smallest possible
piece of context.
Every enterprise has tried the long prompt. It tops out at a few thousand words, goes stale the moment something changes, and still can’t tell the model who the user is or which client they’re writing to. The fix isn’t more prompt. It’s a layered context model that assembles the right facts, from the right sources, for the right task — and cites every one.
That’s what the rest of this page describes.
Layered by scope.
Governed by the people who know.
Company
Mission, brand voice, policies, compliance posture. The things every agent should know before saying a word.
- Tone
- Direct, plain, evidence-led. Never hype.
- Person
- First-person plural (“we”) externally; second-person internally.
- Forbidden
- “unlock”“leverage”“revolutionize”“game-changing”
- Signoffs
- “Best,” (external) · “—” (internal)
- Em dash
- Yes, unspaced. Oxford comma: yes.
# Escalation policy · excerpt
trigger: client mentions litigation, regulator,
breach, injury, data loss
action: route to Legal review
within: 2h
notify: Legal lead, Account lead
# Agents must surface this — never draft a reply directly.Team
How your team operates. Cadences, rituals, artefacts, vocabulary, shared standards.
# team.marketing-ops · vocabulary
our MQLs → leads who booked a demo, not form-fills
our “brief” → a one-pager + stakeholder list + acceptance criteria
our sprint → two weeks, Wed → Tue, retro on the Tue
our review → author + one peer + one outsider; written, not meeting
signoff→ “—[Name], Marketing Ops”
# Agents map generic terms to these before writing anything.- Sprint
- 2 weeks · Wed → Tue · retro Tuesday PM
- Standup
- Async in Slack #mkt-ops-standup · before 10:00
- Weekly review
- Friday 15:00 · 30 min · minutes in Notion
- Planning
- First day of sprint · 60 min
- Default
- Written, async, linkable
- Slack DMs
- Only for people-stuff · everything else #public
- Meetings
- Only if a decision can’t be made in writing
- Docs
- BLUF first. TL;DR last. No walls of text.
Client
Everything known about an external counterparty — history, preferences, contracts, people, tone.
- Account lead
- {account lead} · VP Marketing Ops
- Engagement
- 3 active projects · last touch 2d ago
- Tone
- Formal, cautious, compliance-first. No superlatives.
- Red lines
- No personal data in promptsEU region onlyLegal review for any public statement
- Key people
- {account lead} · {legal contact} · {procurement}
- Commercial
- MSA signed · SOW per project · NET-45
- Gaps
- ⌀ Missing: current OKRs · Q2 roadmap · M&A posture
Project
Goals, decisions, open questions, stakeholders. The story-so-far behind any piece of work.
# project.brief · excerpt
goal : Refresh the {counterparty} SMB brand for 2025 launch
success : 30% lift in unaided awareness · Q3 survey
constraints: Legal sign-off on every public asset
EU data residency · no US-only tools
team : {lead} · 3 designers · 1 copy
# Decisions log
D-04: drop serif subhead · agreed in review
D-05: no AI imagery in launch comms · legal hold
D-06: Swedish first, English second · per client
D-07: soft launch Apr 14 · full launch Apr 28- Q-11
- Do we include the retail wordmark in launch assets? → owner: project lead · needs client call
- Q-12
- Final typography: Tiempos or NHU Display? → designer review Friday
- Q-13
- Approval chain for the press pack? → legal to confirm
- Mar 01
- Brief locked ✓
- Mar 22
- Concept review ✓
- Apr 05
- Legal pre-read ✓
- Apr 12
- Client review ✓
- Apr 14
- Soft launch · this week
- Apr 28
- Full launch
User
Role, responsibilities, writing style, signature preferences, communication tone, recurring tasks.
- Role
- Senior Brand Director · Marketing · Stockholm
- Reports to
- VP Brand
- Voice
- Concise, direct, evidence-led. Rhetorical questions off.
- Default tone
- Warm-professional externally · dry-direct internally
- Signoff
- “Best, {name}” (external) · “—{initial}” (internal)
- Prefers
- Short emails · bulleted when there’s a list · BLUF first
- Avoid
- Emojis in client comms · exclamation points · “just” & “super”
# sample.sent · to counterparty
Hi {counterparty},
Quick update on the launch assets.
- Press pack: drafted, with legal Thursday.
- Film: two cuts, 30s and 60s, previewing Fri.
- Socials: paused until legal clears the pack.
Flagging one thing: the retail wordmark in frame 12
— can we confirm we’re keeping it? Happy to drop if not.
Best,
{name}Every run assembles
only the layers it needs.
“Draft a reply to the counterparty.”
A user asks an agent to do something account-owned. The agent knows which layers are in scope.
- agent: Email Drafter
- user: {operator}
- thread: launch
4 of 5 layers active
Company, team, client, and user contexts attach. Project context skipped — this thread isn’t tied to one.
- ✓ company · voice + policy
- ✓ team · marketing-ops
- ✓ client · {counterparty} (subset)
- — project · not in scope
- ✓ user · {operator}
Drafted + cited
Every claim links to its source in the context. The draft sits in the operator’s inbox, not in the thread.
- Reply drafted
- Citations · 4 layers
- Voice match gate
- Quality gate: ✓ passed
Citations, not opinions. Every agent output comes with the layers and entries that shaped it. Open a draft and you can see exactly which sentence came from which source — and when it was last verified.
Authority plus freshness.
Agents read both before they write.
The team’s source of truth
Brand voice, escalation policy, locked commercial terms. Written by the layer’s owner and signed off; agents weight canonical entries above everything else.
Day-to-day working context
Team rituals, project constraints, account history, user preferences. The bulk of every workspace. Reviewed on the layer’s own cadence, not every week.
Evidence, not verdict
Agent-proposed entries and raw samples — sent emails, call transcripts, wiki chunks. Agents cite them, but canonical and standard entries override them.
Every field has a last-updated
Each field shows an Empty / Thin / Sharp quality dot and a freshness timestamp. When a field goes stale, agents surface the gap in the run instead of guessing from old data.
A dashboard,
not a log.
Rollup per layer. Drill into a layer to see the field list sorted by health, or open a specific entry to see the per-field quality dot and its freshness stamp.
- Quality
- Thin · 42% scored by Haiku
- Freshness
- Last updated 42 days ago
- Authority
- Supplemental
- Owner
- Account lead
Thin fields get assisted, not guessed. Every field has a Haiku-scored quality dot and a freshness stamp. Where a field is thin, “Assist” proposes a fill from your existing artefacts — a human approves, edits, or rejects. The per-agent readiness table shows which agents are context-ready and which are blocked on which fields.
The opinions behind
the architecture.
Context is infrastructure, not a prompt.
It has a schema, a review process, an audit trail, and a health metric. Treat it like a production database — because it is one.
The best context is curated, not crawled.
A scraped Notion page will never beat three sentences written by the account lead. We make human curation cheap, not optional.
Layers scope who owns what.
Brand doesn’t own the client record; the account lead doesn’t own the brand voice. The layer model matches how teams already work.
Cite every claim, every time.
No unsourced sentence ships. If a model asserts something, the entry that produced it is one click away. Auditability is the floor, not a feature.
Staleness is a failure mode.
Out-of-date context is worse than no context. Every entry has a freshness window; past it, agents surface a gap rather than guess.
Context belongs to the team.
Not to the vendor, not to the model, not to the individual. Per-tenant storage, per-tenant embeddings, no training on your data — ever.
The questions teams ask
in the first conversation.
Everyone’s racing to the next model. I’m betting the bottleneck has already moved — it’s not capability, it’s the shape of what you hand the model. Company voice, team rituals, client history, project constraints. Five layers that are almost never written down anywhere an agent can read. If Sempleo is right about anything, it’s that curating those layers is the real work — and the model you use is an implementation detail.
