Narrow Reading Beat Broad Scanning Every Day Of The Week

A Week Of Observation Turned Reading Into A Measurable Craft

This study began with a simple suspicion: the way an agent reads files matters more than the way it writes them. Over five working days I attached an observer to a coding agent and recorded every file it opened, every file it skipped, and the order in which curiosity arrived. The repository was a real product with tests, documentation, dependency declarations, and the usual accumulation of dead branches. I gave the agent ordinary tasks and let its own habits surface. To judge the results fairly I wrote the criteria before the week began. A sound reading habit opens the smallest set of files that answers the task, reaches for entry points before implementation details, prefers the current source of truth over stale copies, and avoids reopening a file it has already understood. Those four rules gave me a scoreboard I could apply to each session without hindsight. The work ran through BestRelay, a relay that forwards each request to the official model named in it, with no substitution, so the habits I recorded belonged to one consistent model rather than to a mixture. Billing there is pay-as-you-go from the token counts the provider reports, with cache reads and cache writes priced separately, which made the cost of each reading decision visible in a way a flat subscription never would.

The First Three Files Shaped Every Later Selection

Across five days the pattern appeared before the first hour was out on each one. When a session opened with a dependency manifest, a test suite, and one high-level module, the reads that followed were short, targeted, and rarely regretted. When a session opened with a large implementation file in the middle of a call chain, the agent drifted, opened neighbours to make sense of the first, and lost time reconstructing intent it could have found near the top. The ordering effect is the finding I would defend hardest: the first three files act like a lens, and every later choice is refraction through them. A session that began near the entry point spent its budget on the task at hand. A session that began mid-machinery spent its budget on geography. The model behind the week was Claude Sonnet 5.5, with a lighter comparison pass on Claude Haiku 4.5, and both showed the same ordering bias, which suggests the effect sits in the shape of the task rather than in any single tuning choice. I also noticed that a good opening set stayed small. Three files seemed to be a natural budget: enough to establish vocabulary and boundaries, few enough that attention remained available for the change itself.

Explicit Criteria Exposed Where Reading Discipline Broke Down

Applying the criteria side by side with the logs made the failures legible rather than impressionistic. The first failure was repetition. The agent reopened files it had read in full minutes earlier, usually while assembling a summary or naming a test, as if freshness of phrasing required freshness of source. The second failure was trust in prose. Old comments and helper notes were accepted over newer declarations, and the agent only recovered when a later compile error forced a second look. The third failure was blindness to cheap answers. Error text and test expectations were underused even though they often settled a question in a glance, while the agent chose instead to read an entire module to infer the same fact. The fourth failure was partial reading that satisfied the smallest set rule on paper while missing the single line that changed behaviour. I scored every session on four axes, namely coverage, order, freshness, and re-read cost. Coverage was strong when tasks were narrow and weak when tasks spanned modules. Order was the most variable axis by far. Freshness was reliable inside code and poor inside prose. Re-read cost was the axis with the most room for improvement, and it is the axis that translates directly into money.

Cache Pricing Made Cheap Reading Look Expensive In Totals

Because the relay bills from the token counts the provider reports and prices cache reads and cache writes separately, I could put a number on each habit instead of arguing about it. Reading a long file for the first time in a session pays a cache write, and that write is charged per token whether or not the file proves relevant. Re-reading stable material afterwards is comparatively forgiving, so repetition was less costly than carelessness. The genuinely expensive habits were different ones: opening broad files where a narrow declaration would have answered the question, and touching many modules after every edit, which invalidates cached content and forces fresh writes across the whole set. Sessions that read narrowly first therefore produced lower totals for the same amount of useful work, and the gap widened as tasks grew. The same relay serves Claude Code, Codex, Gemini CLI, and any OpenAI SDK, so the economics hold wherever the session runs. There is also a referral programme that pays members a commission when people they invite top up, which is the sort of detail worth knowing only after a week of watching the bill respond to reading discipline.

Narrow First And Breadth Later Won Every Comparison

The side-by-side reading finally comes down to a judgement, and mine is that order outweighs volume. Sessions that read narrowly first and expanded only when the task demanded it finished sooner, wrote fewer lines that were later deleted, and produced change sets that reviewers accepted with less discussion. Sessions that scanned broadly at the start looked thorough in the log and delivered patchier work, because the early breadth diluted the attention available for the exact line that mattered. The habit worth teaching is straightforward to state and demanding to practise. Begin at the boundary of the change. Read the declaration and the test before the implementation. Touch only what the task names. Let the suite tell you if your map was incomplete, and let a single targeted read repair it. That sequence respects both the attention budget and the cache budget at once, which is rare alignment in engineering work. The lesson travels beyond one tool or one week: the reading budget is the real budget, and every generous opening scan is borrowed from the precision that follows. Create a BestRelay account and watch your own week of reading with the same scoreboard.