A repeatable method for catching buried guardrails — run today on your own content/ harness.
Why this, for you: this is Lessons 01–02 turned into a tool you keep. You author this harness;
the method below is one you can re-run on any CLAUDE.md / AGENTS.md you write or review,
and one you could publish.
Knowing the U-shape is fluency. Auditing a real file with it is the skill. Here's the
four-step pass — it takes about two minutes per file.
Map sections to attention zones. Mark the top ~20% (primacy) and bottom ~20% (recency) as high-attention; everything between is the dead middle.
Find the hard guardrails. Every never / always / must rule whose violation is expensive. Where does each one sit?
Inspect the tail. What occupies the closing slot — a critical guardrail, or a footer (links, memory notes, changelog)? A footer in the tail is wasted attention.
Measure length. The longer the always-loaded file, the larger its dead middle. Can anything move to a path-scoped rule that loads only when relevant?
Worked example: your content/ harness
CLAUDE.md (~136 lines)
tail-endRead @AGENTS.md + pointer-map✓ primacy used well — costliest-to-violate routing rule first
middleCommit rules: no Co-Authored-By, rebase-before-commit✗ hard guardrails sitting in the dead middle
tail§ Memory hygiene✗ recency slot spent on a moderate-stakes section
AGENTS.md (~87 lines)
start"Two repos — do not confuse them"✓ high-error-risk rule in the primacy slot
middle🚫 Never list · Lethal-trifecta posture✗ the most safety-critical rules in low-attention zone
tail§ Related (two links)✗ near-zero-stakes footer in prime real estate
The pattern is consistent and easy to miss because the files read well top-to-bottom: primacy is sharp,
but both end on a footer, leaving the recency slot — the second-strongest attention zone — unused for
anything that must be obeyed. Meanwhile the hardest never rules sit mid-file.
The fix is a tail, not a rewrite
Don't reorganise 136 lines — Lesson 01 warns that more text means a bigger dead middle. Add a short closing block
that restates only the 2–3 hardest guardrails, so the recency slot finally carries a rule:
# …end of CLAUDE.md…## Critical rules (read last)
- Never add the `Co-Authored-By: Claude` trailer (a hook blocks it).
- Never push, rebase, or delete `production-deploy`.
- `.claude/settings.json` is the single source of truth for hooks + deny list.
↪ Your win: claim the tail in both files
Add a "Critical rules (read last)" block to CLAUDE.md and AGENTS.md — restate, don't relocate.
Pick the 2–3 highest-cost-to-violate rules per file. That selection is an authoring judgment — your call, not mechanical.
Keep it tiny: a restatement that doubles the file would trade a tail problem for a middle problem.
Want me to draft both blocks and show you the diff? Say so — and tell me which guardrails you'd put in the tail.
Retrieval practice — recall, don't peek
Question 1The closing slot of an instruction file should hold…
Question 2In the audited harness, the recency slot was spent on…
Question 3A longer always-loaded instruction file is worse because it…
Question 4 · spaced recall from Lesson 02The U-shaped attention curve is best explained by…
Ask me anything. Want the actual diff against content/CLAUDE.md and
content/AGENTS.md? Or to extend the audit to .claude/AGENTS.md and the sub-agent files?
Say the word — or roll to Lesson 04 on the five per-turn moves.