HomeContext Engineering
Context engineering

Three layers of context around the model.

How an AI agent turns a brief into reviewed work: a canon that humans own, skills that teach the agent how to use it, and live data from the systems you already run. The model is the smallest part. This is the pattern I build inside client stacks, drawn here with the identifiers removed.

Use the template on GitHub Play the loop
Show it for

Marketing: a campaign brief becomes a scored list, content, nurture and a rep handoff. This is the build the page was drawn from.

Figure 1

Brief in, campaign out.

Read it left to right. A campaign brief and a frontier model enter the harness. The harness wraps the model in the team's skills, the team's canon, and the team's live account data. What comes out passes a named reviewer before any tool acts on it. Click a layer to see what is in it and how often it refreshes.

In The harness: your context around the model Out Campaign brief From the growth marketing team market · product · segment offer · timing · channels named reviewer Pilot: one market, one product line (new logo and retention segments) Frontier models Commodity. Swap without rebuilding. Claude · OpenAI · Google hosted workspace or API or inside a campaign tool's harness Layer 3 · Live context (integrations) Real time, through a governed gateway CRM accounts + contacts · live intent signals · live nurture engagement web analytics call recordings, PII removed Partitioned by market. Nothing cached as truth. Layer 1 · Canon Human-owned truth, market-first 6 personas messaging briefs feature base brand voice Layer 2 · Skills (agent harness) brief → context pack routing rules scoring rubric source-system skill reviewer gate lineage Model hosted in a company workspace or inside a campaign tool brief in model in review named human reviewer before anything goes out campaign out Scored target list score · segment · stage · reason to a CRM view, a wiki page or chat (the team's choice) Pages, ads, content a content tool runs the brief landing pages · LinkedIn · Meta Google · personalized pages Nurture the nurture platform runs it nurture emails · landing pages writes activity to the CRM Rep handoff when an account is ready Slack alert · CRM task with the reason attached

What the picture claims: the model is the smallest part. The value is the three layers the team owns around it. Click a layer to see what is in it and how often it refreshes.

Layer 1

Canon
What
  • Product and feature base, pulled from the public support portal
  • Positioning, benefit statements, strategic messaging briefs (jobs to be done, market context, key messages)
  • Persona cards and the ICP for the pilot market
  • Competitive context, including partner rules (who you never attack)
  • Brand voice and imagery rules, corporate narrative
  • Glossary and operating rules
Shape
Market-first. The feature base is canonical; messaging, personas and channel performance sit under each market. One market is the first slice; the others follow using the same frame.
Lives in
One company-owned home: a wiki, a shared drive or a repo. Every page has an owner role, a review date and a status.
Refreshed
Feature baseweekly, automated, owner approves
Messaging briefsper campaign
Personasquarterly or on market change
Brand, commson change
Owner
Product marketing by market, brand and creative, corporate communications. Each owns its slice; off-brand output traces back to the owner, not the user.

Layer 2

Skills (agent harness)
What
  • Brief-to-context-pack: assembles everything one campaign needs from canon and live data
  • Routing rules: market, product, persona
  • Scoring rubric and the product × market × stage campaign grid, encoded
  • Source-system skill: which CRM fields to trust, when to use an export, an API or an MCP tool, how to keep token spend low
  • Reviewer gate, lineage (why an account scored as it did), memory (decisions and edge-case log)
Lives in
A repository under a company-issued identity, loaded into the model workspace project. Any campaign tool loads the same skills into its harness when a brief is run.
Refreshed
Skillseach review loop
Rubricon product marketing sign-off

Outputs are reviewed, the skill is adjusted, and the loop repeats until outputs pass evaluation more than 90% of the time.

Owner
The GTM engineer for the technical skills; product marketing for the content-facing skills. The forward deployed engineer builds the first version and hands over.

Layer 3

Live context (integrations)
Status
Live via governed gateway Agents read the stack in real time through a governed gateway, on every brief. Nothing is exported or snapshotted, so there is no copy to go stale. Swap a connection and nothing in Layers 1 and 2 is rebuilt.
What
  • CRM accounts and contacts, read live through the gateway and partitioned by market
  • Intent signals as they land: board-meeting and RFP signals, champion job changes
  • Nurture engagement and web analytics, including AI-referred and personalized-page visits
  • Call recordings with PII removed
  • Write-back of campaign activity to the CRM through the nurture platform or a custom API app
Refreshed
Real timeread through the gateway on every run, never cached as truth
Owner
The data lead and the GTM engineer own the connections and the gateway. The forward deployed engineer writes the data contracts the skills read against.

Out

Review, then act
Gate
A named human reviewer approves each segment and each output. No segment is activated without review; no spend moves on an unreviewed output.
Surfaces
  • Scored target list with score, segment, stage, recommended campaign and a plain-language reason
  • Content tool: pages, ads and content built from the context pack
  • Nurture platform: emails and landing pages; activity written to the CRM
  • Rep handoff by Slack or CRM task, reason attached
Tools
Interchangeable. The context layers are the asset; a tool reads them, it does not own them.
Figure 2

How context flows and stays current.

The same three layers, drawn as a loop. Owners write canon. A gate publishes it. Skills read it per brief, join live data, and send outputs through the reviewer. Gaps and results come back to the owners. Agents never edit canon. Press Play to walk the nine steps.

1Owners update their slice of canon when something changes. About 20 minutes per page per month.
Canon owners product marketing by market brand and creative corporate communications edit · verify · approve Support portal pull features and functions, public docs agent proposes, owner approves weekly · automated Canon home one company-owned place pages with owner, review date, status market-first, feature base canonical wiki, shared drive or repo Validate and publish contract check: owner, date, status, no secrets owner approves the release gate · nightly sync Read-only bundle what agents actually read versioned, cited by source freshness checked daily repo + model workspace project Campaign brief market, product, segment, offer timing, named reviewer per campaign Skills (harness) brief → context pack routing · rubric · source-system lineage · memory reads the bundle first, live data for live facts Live context CRM accounts and contacts intent, nurture, web analytics call recordings, PII removed partitioned by market live · governed gateway, real time review named reviewer Delivery surfaces not canon; carries provenance scored target list content tool: pages, ads, content nurture platform: emails, pages Slack or CRM: rep handoff CRM view · wiki · chat (list home is the team's call) Gap tracker + edge cases missing or conflicting facts accounts where scoring misfired agents report, never edit canon Performance readback results by segment and channel by market (email differs by region) nurture · analytics · live on change proposes validate nightly reads first per brief real time (gateway) approved only reports gaps and edge cases owner triage within 2 weeks results channel performance by market, back into canon

What the picture claims: freshness is a loop with owners at both ends. Tools read from the bundle; they never become the source of truth.

The ideas

Eight things the pictures are saying.

How I explain the pattern to a leadership team, in the order that lands.

01Brief in, campaign out

The team feeds a brief in; the harness loads the latest context and executes. That is the whole mechanism, and it is what Figure 1 draws.

02Three layers, all owned by the company

Canon is the product, persona and messaging material, organized market-first with the feature set as the base. Skills are how an agent uses it: positioning, routing, scoring. Live context is the data. Campaign tools consume these layers; they are not homes for them. Swap a tool and nothing in the layers is rebuilt.

03Layers 1 and 2 are buildable now

Most product marketing material is in better shape than the team thinks. Once a shared home and accounts exist, the canon, the skills and the refresh automation go up fast.

04Layer 3 is live, never a copy

Agents read account data, intent and engagement at run time through a governed gateway. No exports and no snapshots that go stale between campaigns. Access is read-only and governed per skill, so the data team keeps control of what each agent can see.

05The connection can change without rebuilding anything

A warehouse partitioned by market, a governed gateway for agents, custom API apps instead of metered gateway calls. Swap any of them and the same skills keep reading live data. Nothing in Layers 1 and 2 moves.

06Freshness is owned, not automatic

Feature canon refreshes weekly from the support portal with owner approval. Messaging, personas, brand and comms change on their owner's cadence. Off-brand output traces to the owner of the stale slice, not to the person who ran the brief.

07Brand tuning is a loop

More examples and templates, review the output, adjust the skill, repeat until it passes more than 90% of the time. That is the training work with the creative team, and it is the same loop an AI Pod runs.

08Review before anything goes out

A named reviewer sits between the harness and every tool. Nothing is activated or spent without it. Autonomy is earned later, one reviewed output at a time.

Use it yourself

The governance behind this page is open source: the canon contract, the gap tracker, the gated sync, the metrics catalog, and a hub-and-spoke variant for consultancies. Copy the repo, rename the slices to match your business, and start with Layer 1.

Get the template on GitHub

Drawn from a real build for a B2B marketing organization, with client identifiers removed. The other six departments are the same pattern mapped onto their own canon, skills and systems of record. The repo at github.com/justin-gpt/context-engineering-template is the generic version.

FAQ

Context layers, answered.

The questions that come up when a team sees the picture for the first time.

What is a context layer?

A body of information that sits around the model and is loaded at run time. There are three: canon, the human-owned truth about the business; skills, which teach the agent how to use that truth; and live context, current data from the systems the team already runs. The model is swappable. The layers are the asset.

Why separate canon from skills?

Canon is what is true about the business and changes on an owner's cadence. Skills are how an agent uses it: routing, scoring, assembling a context pack, gating output. Keeping them apart means a messaging change never requires touching the agent, and an agent change never risks rewriting the truth.

Why do agents never edit canon?

Because provenance would break. Agents report gaps and edge cases to a tracker. A human owner triages and edits, so every page keeps an owner, a review date and a status, and a correction made once stays corrected.

What keeps the live layer from going stale?

Nothing is copied. Layer 3 is read through a governed gateway at run time, on every brief, so the agent sees the current state of the systems of record, whichever department it serves. If a connection changes, the gateway is the only thing that moves. Nothing in the canon or the skills is rebuilt.

Where does the human reviewer sit?

Between the harness and every tool. A named person approves each list and each output before anything is activated, sent or spent. The tools downstream are interchangeable; the gate is not.

Does this only work for marketing?

No. The page was drawn from a marketing build, and the department switch at the top maps the same three layers onto sales, product, customer success, operations, engineering and HR. What changes is what each layer holds and which systems the live layer reads. The pattern, the reviewer gate and the loop stay the same.

Want these layers around your model?

Book a Call Get the template on GitHub