Build1 publisher3 min readPublished
Seven months of building grew one pipeline's agent guidance to 1,042 lines
Newer models ship with advice to keep the guidance file under 200 lines. The analyst who wrote those 1,042 lines argues each one logs context the model lacked, and that the real debt is a mechanical rule left sitting in prose.
The Engineer · Build desk
What happened
- A two-person content pipeline built from February 2026 on Claude Opus 4.6 and Sonnet 4.6 now carries 1,042 lines of guidance prose across three files, one of which is global to all the analyst's projects.
- Alongside that prose the same project runs three mechanical gates, none of them a test, and seven content validators.
- The first CLAUDE.md, dated 9 February 2026, was 54 lines: a two-sentence overview, three commands and seven conventions.
- The first pipeline run, from 4 February to 10 March 2026, produced 222 commits, 89 labelled features and 82 labelled fixes.
- The team dates itself on Boris Cherny's five adoption steps as gated in January, parallel from 22 February, and at supervised autonomy since June without reaching step 4.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Trimming to a target turns into a per-line decision with three answers: leave it in prose, move it into a hook that fires on the tool call, or accept it as a fact the code cannot supply.
- constraint A budget stated as a line count treats a redundant line and a check nobody remembers writing the same way, so cutting to hit the number risks removing the check with the line.
- contradiction The tool's makers advise cutting the file; the practitioner keeping 1,042 lines argues the growth records context the model lacked, and the supplied account does not classify a single line either way.
Each line in that file carries a cost. It enters the session with everything else, and the analyst's own accounting is that you pay for each line in context, in each session, whether or not the model reads it that day [7]. A hook works differently. In Claude Code a lifecycle hook fires before or after a tool call [9], so a rule expressed there costs nothing to carry and does not depend on the model agreeing with it. The harness in this project is all three layers at once: the CLAUDE.md the model reads, the hooks around its tool calls, and the validators over its output [8].
The sort therefore has three bins, and the analyst names them: lines the model now carries by itself, and lines that belong in a hook. The third is facts no model will infer from the code, however good it gets [6]. "Growing instructions for the model is not debt. It is the record of each piece of context the model lacked," the analyst wrote in a dev.to account of the project. "The debt is the shape: a mechanical rule that stayed in the prose the model reads instead of the code that stops it" [5].
Set against the under-200-lines advice [1], the guidance prose runs about 5.2 times over [1], and about nineteen times the size of the 54-line first file [2]. One of the three files is global to all the analyst's projects [3], so not all of that total is charged to this repository. The deterministic side of the same harness holds ten checks in total [3].
For 200 lines to be the right budget in a given repository, either the guidance would have to be free of the third bin, or the deterministic layer would already have to carry the mechanical rules. The third bin here is easy to point at. The plan merged on 31 January 2026 covered site structure, personas, tone of voice and the conditions for fact-checking [13], and the repository started as a Hugo site with a CMS [12].
The post gives the dated record and the shape of the question, but does not break the 1,042 lines down by bin. Two cautions apply to the commit figures. The analyst says the 222 commits from 4 February to 10 March are counted by label, so a one-character correction and a rewritten step count the same [15]. The 89 features and 82 fixes leave 51 commits carrying neither label [4]. The ladder under all of it is Boris Cherny's, described on 16 July 2026 [10]. The analyst concedes it is a vendor's frame, and keeps from it only the bottleneck at each rung and the thing that unlocked the next [17].
Step 4 on that ladder assumes a thousand agents or more [10]. This pipeline is two people who had not built with agents before February 2026 [2], and it has sat at supervised autonomy since June [11].
What to watch
- Whether the project publishes the per-bin classification of the 1,042 lines, and what the three gates and seven validators actually check.
- Whether the makers of the tool attach a workload to the under-200-lines guidance or leave it as a bare count.
- Whether a team stuck at supervised autonomy since June names what would unlock Cherny's step 4.