System instructions
The standing rules: voice, refusal, format.
AI Training
Give AI the right information.
most AI failures aren’t bad wording — they’re missing, stale or wrong information. Context engineering fixes the inputs.
Why context beats prompting.
same prompt, different context, different result. Models can only use what they can see; a perfect prompt over stale data produces confident nonsense.
The four layers.
Project knowledge
Your documents, specs, brand book — curated, dated.
Retrieved context
The right excerpt for this task, with provenance.
Task input
The actual work item.
A context manifest.
every retrieved document gets a manifest entry so provenance is checkable.
Context Manifest Example
# sources: product-spec v4 (2026-09-30, owner: A. Reyes) # freshness: refresh on every product release # claims: only from source · flag anything undated # size: ≤ 8k tokens · dedupe · no PII, no secrets 1sources: product-spec v4 (2026-09-30, owner: A. Reyes) 2freshness: refresh on every product release 3claims: only from source · flag anything undated 4size: ≤ 8k tokens · dedupe · no PII, no secrets
Context hygiene.
- dated sources or they don’t ship
- dedupe before the model sees it
- provenance on every excerpt
- size budget per task
- protected values listed
- no secrets or customer data in prompts
Ground it, then verify.
-
Audit
what the model sees today, item by item
-
Curate
kill stale, duplicate, unowned content
-
Ground
retrieve with provenance, manifest every source
-
Verify
claims-check output against the manifest
“Ready to fix the inputs?”
Start with your situation. We’ll audit what your AI actually sees.