F2 Moving what only some turns need out of the every-turn file into trigger-loaded files cuts the every-turn cost by a measured amount

Field	Value
Status	active
Topic	Context economy and agent prose
Grade	R3/S3
Cases	Shio, roadkeep, winwright
Contrary	none
Harness facts	H1, H2, H3

Claim

When content needed only on some turns was moved from the file loaded on every turn into skills or files read on demand, the every-turn file shrank by a counted amount, and the projects kept the moved content out of the every-turn file by rule.

Mechanism

Everything in CLAUDE.md and what it imports loads every session (H1, H2), while a skill's body loads only when the skill is used and until then costs only its name and description (H3). A rule needed only when building, committing or writing a governed file is therefore paid on every turn if it sits in the every-turn file, and only on the turns that need it if it sits in a skill whose description names those turns.

Observations

Shio. The split of 2026-07-29 [shio@f4ffdb0d9] took agents.md from 185,734 bytes

[shio@e73516a9f] to 14,433 (git cat-file -s), moving the detail into docs/agents/ and eight skills added in the same commit. The index now states the rule it follows: "it is deliberately thin: it is loaded into every session on every turn, so P3 applies to it" [shio@821f18d74:agents.md#L61-L62]. The ignore file names the skills directory the "lazy-loaded half of agents.md" [shio@821f18d74:.gitignore#L72].

roadkeep. RK23 recorded that "Rules resident every turn spend the budget they exist to protect" [roadkeep@91754240:docs/CHANGELOG.md#L970]

and moved the write path "out of the resident file and into a trigger-loaded skill" [roadkeep@24e6fcda]; the budget comment records the reading after it, "120 lines, 7695 bytes once RK23 moved the" [roadkeep@91754240:roadkeep.toml#L238] write path out. RK1136 measured that "26 lines of the every-turn file are needed only on a turn that builds or commits" [roadkeep@91754240:docs/CHANGELOG.md#L667], moved them to a project skill [roadkeep@b12d28b5], and the file "dropped to 104 of 125 lines" [roadkeep@91754240:docs/CHANGELOG.md#L667]. The skill says of itself "Trigger-loaded, and that is the whole reason it is a file (RK23, RK1136)." [roadkeep@91754240:.claude/skills/roadkeep-dev/SKILL.md#L8] The rule that keeps it so is a non-goal: "What loads every turn is only what a turn touching no governed file needs." [roadkeep@91754240:agents.md#L111]

winwright. WW69 was filed as "the skill is loaded on every turn against a budget it does not need" [winwright@861b82e:docs/CHANGELOG.md#L349]

and shipped as "load the skill on a window rather than every turn, and measure both of its costs" [winwright@7c456e0]. The test that holds it asserts that the description names the occasions on which the skill should load [winwright@861b82e:tests/Winwright.Tests/SkillTests.cs#L45-L54]. This concerns the skill winwright ships to adopters, not an instruction file of its own, and no before-and-after figure is recorded.

Independence

Three origins. Shio's split answered Shio's own measured failure (F1). roadkeep's RK23 came a day later and may share the idea with Shio's split; it counts as a second origin because roadkeep recorded its own need, its own file at its ceiling and 26 lines of it used only on build or commit turns (grading rule 2, a failure of its own that the remedy addressed). winwright's WW69 answers a loading problem in its own shipped skill. freewilly and winwright carry no every-turn file at all and keep their working rules in skills; no source at either pin says why, so that arrangement is not counted here.

Counter-evidence

The move reduced the every-turn file, not the total. At the split, docs/agents/ held

184,521 bytes and the eight new skills 32,867 (git ls-tree -r -l f4ffdb0d9), together more than the 185,734 bytes they replaced. Whether a task then read less in total is not known: session transcripts are not a source in this study. The growth of the moved content is F3.

winwright's observation is at S2: the ledger records the change, but no size is counted.

Grade

R3: the move and its effect are observed in three projects, each answering its own problem. S3: Shio's byte sizes before and after the split, and roadkeep's line counts of 104 of 125 after RK1136, are counted. The per-turn saving itself is documented harness behaviour (H3), not a corpus measurement.