Context budget discipline

The discipline the field has converged on has four parts. Keep the always-loaded layer thin, and make it point rather than contain. Prefer the smallest context that completes the task over the most context that fits. Pass references to durable objects instead of copying their content into the window. And treat the tool surface itself as context, since every tool description is charged to the same budget.

Every group that measured it found accuracy improved, or held flat, while tokens fell. The wins do not come from spending less; they come from removing material the model was being distracted by. That reframes context trimming from a finance exercise into a correctness one.

Skill and tool descriptions draw on the same window. With dozens of skills competing for a small default slice, whether a skill fires at all becomes an optimization problem separate from whether it works — and most harnesses cap description injection at a percentage of the window and then truncate silently, breaking activation cues. A skill that never fires is invisible to outcome-only measurement.

Also: Headroom’s pip-installable context proxy; Tessl; and Anthropic’s published skill practices (progressive disclosure, concision, explicit activation declaration).

A thin always-loaded layer — short house rules plus the installed role skills — with everything else behind on-demand loads: the methodology spec, the briefs for a stream, and a findings corpus that is cited by link and never restated. That cite-never-restate rule is pass-by-reference discipline, written down before anyone quantified it.

Assay’s always-loaded layer has grown. It carries a placement rule — only rules that bind every session — but no size or density gate, and nothing measures whether a given rule is load-bearing. Two mechanisms are designed and unbuilt: a size and density lint on instruction files, alongside the parity check that already runs over skills; and a standing audit that moves weekly-or-rarer guidance out of the always-loaded layer into hooks, rulesets or CI assertions. That second one is simply placement-beats-content applied to Assay itself. There is likewise no activation telemetry — no count of which skills actually fire — so a skill truncated out of the window would be exactly as invisible here as the evidence describes.

The orthogonality result reads as outside support for keeping the desk verbs narrow and non-overlapping: that is a reliability property with a number behind it, not tidiness. And the split between an always-called setup set and a discretionary set maps onto session boot plus on-demand skills.

• Whether anyone publishes a dose-response curve — output quality against instruction-file size on a fixed task — rather than the current mix of anecdote and cited-but-unnamed studies.
• Whether reversible compression (the re-fetch breadcrumb) holds up outside the vendor’s own benchmark, and whether any harness adopts it natively rather than as a proxy.
• Whether activation becomes a measured property anywhere: per-skill invocation counts in a harness’s own telemetry would move this from advice to instrumentation.