Decision 062: Third-party writing skills are registered, not vendored — and their ban-lists are not adopted

• Status: proposed
• Date: 2026-08-06
• Deciders: @MadaraUchiha-314 (issue #165)
• Work item: issue-165
• Spec: docs/specs/issue-165/ (survey in brainstorm.md)
• Follows: the register-only posture already applied to caveman and ponytail (issue-37), and decision-005's stdlib-first, bundle-no-runtime stance.

Context

Issue #165 asked for a literature survey before building anything. Surveyed August 2026: avoid-ai-writing (MIT), anti-ai-slop-writing (MIT), Stop Slop, several technical-writing-expert skills, the plain-language style-guide tradition (Google / Microsoft / GOV.UK), and Diátaxis.

The finding that shaped the design: every existing skill treats this as a vocabulary problem. Ban "delve", "leverage", "robust"; rewrite the sentence. That is the wrong altitude for the-loop. A 30 KB requirements document does not get better by losing the word "leverage" — it gets better by being 3 KB.

Decision

Register the prior art; adopt its structure and its warning; leave its ban-lists.

• avoid-ai-writing, anti-ai-slop-writing and stop-slop are entries in externalTools.tools with notes recording exactly what was and was not taken. An operator who prefers the packaged version installs it; the-loop bundles no copy.
• Taken: severity tiering (P0/P1/P2) and detect-before-rewrite from avoid-ai-writing; the short-contract-plus-long-catalogue file shape from anti-ai-slop-writing; throat-clearing openers as the highest-yield cut from Stop Slop; front-load-the-conclusion and one-idea-per-paragraph from the style guides; from Diátaxis, the diagnosis that documents bloat when they mix explanation with reference — which is why the-loop's answer is a per-artifact shape rather than a per-sentence rule.
• Not taken: the word-tier ban-lists, the blanket em-dash ban, and the four-mode taxonomy. "Leverage" and "robust" are ordinary technical English; a build that goes red over them is a build people route around. the-loop's artifacts already have names and gates, so they do not need Diátaxis's.
• Taken as a constraint: avoid-ai-writing's own warning that these patterns are signals, not proof, with false-positive rates above 60% on non-native speakers. It is the reason test_writing_parity.py's P4 asserts only tells with no legitimate technical reading, and why the rest of the catalogue stays in skills/writing/reference/tells.md as judgement.

Consequences

• One place to look when a writing rule is questioned: the skill, with the survey behind it in the work item's brainstorm.
• The registry entries carry the reasoning, so a future contributor proposing "let's just install the popular one" finds the answer already written down.
• If the tells catalogue drifts from the upstream skills, nothing breaks — it was never a fork.