Theory and Sources

Four independent lines of evidence converge on the same structure. This page separates what is documented from what is measured locally from what remains assumption, because a design whose provenance is unclear cannot be argued with.

1. Instruction-following degrades with density

Peer-reviewed benchmarks, independent of any vendor's product claims.

IFScale (Jaroslawicz, Whiting, Shah, Maamari, arXiv:2507.11538) evaluates instruction-following at increasing density across twenty frontier models from seven providers. Even the best models fall well short of full compliance at the maximum density tested. Three degradation patterns appear — threshold, linear, exponential — along with a measured bias toward earlier instructions.

SEQUOR (arXiv:2605.06353) finds accuracy on a single constraint declining as a conversation lengthens, and multi-constraint accuracy declining several times faster.

LongGenBench and LIFBench (arXiv:2411.07037) show prompt-adherence loss in long contexts generally.

The design consequence: a rule's position and the total instruction count both affect whether it is followed. A 4,000-word constitution is not four times the governance of a 1,000-word one. It may be less.

2. Vendor guidance changed direction

Anthropic reported removing over 80% of Claude Code's system prompt for its Claude 5 generation models with no measurable loss on internal coding evaluations, and described six shifts: rules to judgment, examples to interface design, upfront context to progressive disclosure, repetition to simple tool descriptions, manual to automatic memory, and simple specs to rich references.

The same source identifies overconstraining across system prompt, rules file, and skills simultaneously, with conflicting instructions inside a single request — the example given being "leave documentation as appropriate" against "DO NOT add comments."

The Opus 5 prompting guide instructs removing explicit verification instructions, because they cause over-verification and waste tokens without quality gain. The Fable 5 guide adds that skills written for earlier models are often too prescriptive and can degrade output.

Claude Code's own documentation states the operational form: a bloated rules file causes Claude to ignore your actual instructions, and the test for each line is whether removing it would cause a mistake.

Treat this as directional, not audited. It is Anthropic's figure, on Anthropic's evaluations, for Anthropic's models, without a per-category breakdown. It is strong evidence about those models and weak evidence about anything else.

3. Advisory versus deterministic

Anthropic's hooks documentation draws the line directly: rules-file instructions are advisory, hooks are deterministic and guarantee the action happens.

This is the load-bearing insight behind moving tier discipline, single-writer builds, the anti-sycophancy gate, and the compaction re-anchor out of prose. Those are predicates, not judgment calls. A predicate checked by a process does not compete for attention with 200 other instructions.

Critic isolation follows the same logic one level up: a subagent that structurally cannot see generation history is a stronger guarantee than an instruction asking a model to ignore what it already knows.

Independent support for external verification comes from multi-agent verification research — BoN-MAV (Lifshitz, McIlraith, Du, arXiv:2502.20379) reports stronger scaling from multiple verifiers than from self-consistency, including weak-to-strong generalization.

Note the tension this resolves. Anthropic says remove verification instructions from prompts; verification research says external verifiers help. Both hold: verification belongs in the loop — tests, stop conditions, hooks, CI — not as prose nagging in a rules file.

4. Compaction determines durability

Content loaded from disk at startup is re-read after compaction and re-injected. Content that arrived through the conversation is summarized, and summarization is lossy. Standing policy is among the first things discarded, because it does not resemble the task in progress.

Corollary with a measurable cost: @import resolves at launch, so an imported file loads in addition to the importing one. A CLAUDE.md containing @AGENTS.md loads AGENTS.md twice. A symlink does not.

Measured locally

These numbers come from running the tooling on real repositories, not from any published source. They are reproducible with the shipped scripts.

Measurement	Value	How
Resident before/after on one repo	9,393 words across two files → 1,396 managed + 866 carried in one	wc -w, verify.sh
v3 rule IDs resident after migration	45 → 0	grep -cE '^\*\*[A-G]-[0-9]+ ·'
Profile sizes	lean 847 words, mixed 1,396	verify.sh
Skill budget on one machine	2,267 skills, ~163,000 tokens vs ~4,000 budget, ~41× over	skill-budget.sh
Waypoint schema across the estate	.phase is an identity string; .status is the lifecycle	jq across every waypoint

That last one produced a real defect. An early tier-guard.sh matched .phase against milestone|release|certify. No waypoint in the estate could ever satisfy it, so Tier 3 was blocked unconditionally with no reachable unblock path. The rule was written from an assumption about a schema; measuring the schema refuted it.

Assumptions, still unmeasured

Stated as assumptions so they can be attacked.

• Per-model scaffold need. The table in Model Profiles gives starting assumptions for Kimi, MiniMax, GPT, and local models. Only the Claude entries rest on published guidance. The rest is reasoning from general principles about smaller models and should be replaced by your own task-set measurement.
• The 1,500-word ceiling. A judgment about where the resident file stops earning its cost, not a measured threshold.
• The migration map. Where each v3 rule lands is a judgment about equivalent coverage, not a verified claim that behavior is preserved.

What would refute this

A design that cannot be refuted is not a design.

Fixed task set, run twice. Ten representative tasks for your repository, run under the old file and the new structure, per model in the fleet. If pass rate regresses at lower token cost, specific deleted rules were load-bearing for your codebase and the reduction was wrong for you.

Compaction test. Force a compaction mid-session and check whether the invariants still hold. If behavior drifts on capability inversion or tier discipline, the invariant was not actually in the durable channel and the layering claim fails.

Hook audit. If a hook is installed but not referenced in settings, enforcement is zero while the prose that asked for the same behavior has been deleted. That window is strictly worse than either state alone. verify.sh checks it; check that verify.sh is being run.

Budget measurement. If skill-budget.sh still reports many multiples over budget after gating, the plugin split was insufficient and the router must carry more.

The strongest case against

Vendor evidence for aggressive reduction is a vendor's own number on its own evaluations. "Trust the model's judgment" lands differently in a regulated, contractually-governed codebase than in a hobby project — implicit judgment is precisely what a reviewer of governed AI does not want.

A verbose constitution that reliably prevents one capability-inversion breach, or one sycophantic sign-off on a client deliverable, may be cheaper than the tokens it costs.

The disciplined answer is not "keep the large file" and not "cut it because a blog post said so." It is: delete nothing until measurement shows it is safe to delete, and treat the migration as evidence-gated. That is why verify.sh fails a repo that declares lean without a measured fleet entry, and why the default profile keeps the scaffold rather than assuming it away.

Source index

Claim	Type
80% system prompt reduction, six context-engineering shifts	vendor, primary
Remove verification instructions on Opus 5; Fable 5 over-prescription	vendor, primary
Bloated rules files cause instructions to be ignored	vendor, primary
Hooks deterministic, rules advisory	vendor, primary
Project-root file re-read after compaction; imports load at launch	vendor, primary
Claude Code does not natively read AGENTS.md	vendor issue tracker, open
skillListingBudgetFraction, silent description drops, eviction ranking	vendor changelog + issue reports
AGENTS.md as an open convention across 20+ tools	open standard, widely adopted
Instruction-following degradation with density and turn count	peer-reviewed (IFScale, SEQUOR, LIFBench)
Multi-verifier scaling	peer-reviewed (BoN-MAV)
Resident word counts, budget ratios, waypoint schema	measured locally, reproducible
Per-model scaffold need outside Claude	assumption, unmeasured

Vendor documentation moves. Where a number here disagrees with current documentation, the documentation wins — and the local measurements are reproducible with the shipped scripts rather than taken on faith.