When Trying to Make AI Better Makes It Worse — Every (Working Overtime, Katie Parrott)
Why this is in the vault
A first-person account of context rot by over-instruction: a working AGENTS.md/style-guide setup degraded over weeks as the author added a durable-record rule, zettelkasten-style cross-links, and ever-more-codified style rules — producing 127 drafts and 1M+ words on a single essay before she diagnosed and reversed it.
The core argument
Katie Parrott, a staff writer at Every, had a working per-column context system: each writing folder held a short voice guide, style guide, and example drafts, read via an AGENTS.md pointer file. It worked well — until she tried to improve it. She made three changes: (1) instructed the model to leave a "durable record" of every experiment, including failed and abandoned ones, with repeated corrections auto-promoted to checklist items; (2) adopted a Zettelkasten-style practice of cross-linking every idea to every related idea, until 385 files pointed at each other; (3) codified every observed pattern into her style/voice guides, growing one guide from 759 to 3,855 words with eight approved openings and six approved endings.
Each change individually seemed sound. Together, they meant the model read an ever-growing, self-contradicting pile before writing a word, and — critically — "couldn't tell a record from a rule, so it followed all of them." One essay draft generated 127 versions and over a million words. Diagnosing it (with help from a multi-reviewer doc-review workflow) surfaced the mechanism: "concrete rules are easier to verify, so they can overwhelm the subtler ones while every checklist still passes." Mechanically checkable rules ("open with one of these eight approaches") crowd out judgment-based ones ("let one experience carry the piece") because the model can tick the former and not the latter.
Her fix was not better rules but fewer of them: she archived the entire tangle (kept, never referenced by default), rebuilt from an empty folder with the model interviewing her about her actual needs, and landed on three principles — the agent asks before writing to any guide file (nothing becomes an instruction without her approval); every folder/task is self-contained and doesn't follow cross-links into unrelated projects; and guides explicitly mark drafts/logs/notes as "task material or evidence, never new instructions." The rewritten guide ran 513 words, down from 6,109. She adds a personal rule on top: no major context-architecture decisions under pressure (deadlines, bad weeks, or new-model excitement) — write the idea down and revisit it later instead of acting immediately.
Mapping against Ray Data Co
This is a strong, concrete match — it's a lived post-mortem of exactly the failure mode AGENTS.md's own two-tier structure ("Hard rules (mechanical, no exceptions)" vs. everything else) was built to avoid, and it's a live warning about this vault's own MEMORY.md. MEMORY.md here has grown to ~90 feedback/project/reference entries, almost all originating the same way Parrott's style-guide bloat did: a one-off correction ("founder caught X") gets promoted into a standing memory line that every future session must weigh against every other line. Parrott's diagnosis — "the model couldn't tell a record from a rule, so it followed all of them" — is a precise description of what the AGENTS.md precedence chain exists to prevent (rule #2, "memory conflict → newer wins, and log it... don't silently apply"), but the chain only resolves conflicts it can see; it doesn't stop accumulation, and Parrott's essay shows accumulation alone degrades output even without outright contradiction.
Her second finding — that mechanically-checkable rules silently crowd out judgment calls because "every checklist still passes" — is a sharper version of the concern behind the auto-mode classifier hard-gate memory and the verify-* critic fleet: a producer satisfying every explicit rule can still miss the thing the rule was a proxy for, which is exactly why RDCO routes vault writes, dispatches, and strategic outputs through a fresh-eyes critic rather than trusting the producer's own checklist compliance.
Her rebuilt-system principles map almost one-to-one onto RDCO's existing practice, which is worth naming as validation rather than new insight: "agent asks before writing to a guide file" is the spirit behind PR-only workflow and founder-gated writes; "every folder stands alone, doesn't follow cross-links by default" mirrors the lazy-pointer design of CAPABILITIES.md (read on-demand, not loaded every session); and "drafts/logs are evidence, never instructions" is the same distinction AGENTS.md draws between hard rules and narrative/working-context content that should never drift into rule status.
The one place this essay doesn't map is tactical: Parrott's setup is single-human, single-model, column-scoped — there's no equivalent here of a cron-driven always-on agent accumulating state across unattended runs, which is where RDCO's actual rule-pile-up risk differs from hers (hers comes from one writer's accumulated preferences; RDCO's would come from skills/memories self-amending without a human noticing). That's a gap worth flagging rather than forcing a fit.
Related
- [[2026-10-02-every-codex-graded-my-ai-habits-then-coach]] — same Every "Working Overtime"-adjacent personal-AI-habit genre, but the opposite lesson: that essay is about building a recurring self-correction loop; this one is about what happens when self-correction loops are left to accumulate unchecked. Read together as a matched pair.
- [[2026-09-24-every-why-evals-are-so-hot-right-now]] — companion Every argument that personal benchmarks/context should be built from a person's actual recurring work rather than theorized in advance, the same anti-over-engineering instinct Parrott lands on.
- [[2026-07-26-harness-seven-failure-mode-scorecard]] — RDCO's own harness self-audit against failure modes; rule-pile-up/context-rot from unchecked accumulation is a failure mode this scorecard doesn't yet name explicitly and could.
- [[feedback_no_claudemd_state_drift]] — the existing house rule ("no session state in AGENTS.md; stable rules only") that this essay validates from an independent, non-RDCO source.
⚠️ Sponsorship
House self-promo only: the footer pitches Every's own "All Access" bundle (Sparkle, Cora, Spiral, Monologue) and a subscribe CTA, standard for every Every issue. No third-party paid sponsor block in this issue. Bias implication is low — the essay's substance (a personal, warts-and-all narrative about the author's own workflow failure) isn't shaped by the self-promo, which is confined to the footer and unrelated to the tools discussed in the body (Claude, ChatGPT, Codex/GPT-5.6, Every's own Compound Engineering doc-review plugin — the latter is a genuine in-body Every-product mention, used diegetically to diagnose the problem rather than pitched).