Context is a range, not a floor — prune past the sweet spot and the agent points the wrong direction just as confidently as when you bloat past it.

Principle

Context Has a Sweet Spot

The shape of the problem

Two cliffs, one plateau.

Context engineering is framed as monotonic. It isn't.

← starve

sweet spot

pollute →

Under-context confident, wrong-directed output

Uniquely specifies minimum inputs, correct frame

Over-context distracted, noisy output

← reduce works only from above the sweet spot

add → works only from below the sweet spot

Context sets the agent's reference frame — both directions shift it off-target.

Why it's dangerous

Over-context fails loudly. Under-context fails silently.

Over-context — visible failure

Signature

· mentions irrelevant files

· hedges more

· quality drop is obvious

Diagnosis time: minutes. Fix is well-understood — prune the noise, rerun.

Under-context — invisible failure

Signature

· confident, coherent output

· looks fine on review

· pointed at the wrong target

Diagnosis time: production. Caught in code review if you're lucky — ships otherwise.

The dangerous failure is the one that looks like success — under-context is confident guessing.

What the sweet spot is

The minimum set that uniquely specifies the output.

sweet spot = minimum inputs such that exactly one correct output exists

1

Just the files to write the solution

A Relevant Files list — not the full repo.

2

Observable acceptance criteria

The constraint that separates right from plausible.

3

The anti-requirements

Explicit what-not-to-do. Bounds the surface.

4

Domain conventions

AGENT_SYSTEM.md, the mental model. The shared vocabulary.

Drop any one and the answer stops being unique — add anything else and the frame gets noisy.

How teams get it wrong

The violations are all one-way thinking.

mono-rule

Treating "less context is always better" as a one-way rule — as if the risk only lives on one side.

over-prune

Pruning workflow-critical documentation to hit a token target. AGENT_SYSTEM.md loses sections and builder agents rediscover conventions per task.

under-spec

Passing a planner an issue body with no codebase references. The output is confident — aligned with something other than what you wanted.

monotonic

Framing context as a floor ("reduce more") rather than a range with two failure modes. Regression that looks like model drift but is self-inflicted.

Every violation here assumes context pressure runs one direction — the principle is that it runs two.

The operational rule

Prune to the first drop — not to a number.

▮▮▮

Start bloated

everything attached

→

▮▮

Prune & measure

watch output quality

→

▮

Stop at first drop

back up one step

Prune until the first drop in quality— not —until token count hits a target

Token-count targets are proxy goals — quality is the real signal, and it turns down before the budget bottoms out.

The discipline

Reduce means reduce from above the sweet spot — never below it.

Pair every prune with a calibration check. Over-context is fixable noise; under-context is confident, invisible wrongness. The craft isn't minimizing — it's staying inside the range.