Recognise a context failure
Keep the instruction good but remove one crucial fact or add a misleading source.
Prove: distinguish prompt failure from context failure.
Keep: context failure audit
Work With AI / field guide / 8 labs
Context Engineering
Give AI the right reality—not simply more tokens, more files or a bigger prompt.
Work With AIContext EngineeringField guide
Work With AI field guide
What you'll learn
Choose the minimum useful context, decide which source should win, expose conflicts and unknowns, package context cleanly and inspect whether the output is actually grounded.
Bring / required
One real task plus two or more possible sources/files if available.
What you'll do
Keep the instruction reasonably good, then deliberately vary the context so you can see when the information—not the prompt—is the failure.
Take away
A context failure audit, source-of-truth map, context pack, freshness rule and grounded-output check.
Use the shared Field Guide Lab Method so context changes are compared against a baseline rather than judged from one good-looking answer.
The fluent-answer problem
A model can follow your instruction beautifully while working from incomplete, stale or contradictory reality.
Aha
A better instruction cannot compensate for the wrong reality.
Common traps
- Context stuffing: every available file goes in.
- Availability bias: the easiest source to attach becomes the source of truth.
- Recency confusion: newer message automatically beats approved policy.
- Silent conflict resolution: the model chooses between sources without surfacing disagreement.
- Attachment theatre: “the file was attached” is treated as proof that the key claim came from it.
Look out for
- Generic answers despite a detailed prompt.
- Fluent claims whose source cannot be located.
- Several documents with different dates or owners.
- Facts that were true last month but may not be true now.
- Important unknowns converted into confident bridge assumptions.
Progressive labs
ENGINEER THE REALITY AVAILABLE TO THE MODEL.
Begin with the work decision. Add only the context that should legitimately shape it.
Find the minimum useful context
Aha: more context can add noise and contradictions.
Prove: remove one source and see whether outcome quality changes.
Keep: minimum-context list
Decide which source wins
Authority · date · scope · specificity · supersession · applicability.
Trap: available = authoritative.
Keep: source-of-truth map
Expose conflicts and unknowns
Aha: “I don't know” can be the highest-quality result.
Prove: conflicting sources must be surfaced, not silently averaged.
Keep: conflict/unknown record
Package context for the job
Selected file · excerpt · structured table · example · current case data.
Trap: attach the whole folder.
Keep: context pack
Give context a validity window
When was it true? For which scope? Has it been superseded?
Aha: context has time.
Keep: freshness / effective-date rule
Separate instruction from context
Keep what to do distinct from what to work from.
Trap: policy, facts, instruction and output template mixed in one block.
Keep: instruction/context pattern
Prove the important claim is grounded
Ask for evidence, conflicts, unknowns and inspect the source of the claim that matters.
Trap: citation-shaped text without actual support.
Keep: grounded-output check
Where Context Engineering stops
Work With AI teaches you to choose and shape the right reality for a task. When you delegate the responsibility, a new problem appears: the human should not need to rebuild that context pack every run.
Boundary
Context Engineering decides what the model should know. Harness Engineering operationalises how a delegated agent reliably finds that context across tools and sessions.