Launching soon on Product Hunt
Claude Opus 4.7 ships with a 1M-token context window. Filling it is easy. Using it well is the discipline. Production-grade strategy for context budgets that produce better diffs.
The 1M-token context window in Claude Opus 4.7 is the single most overhyped and most underused capability in the model. Overhyped because most teams cannot use it well. Underused because the teams who can use it well still rarely do.
This is the strategy BridgeMind runs internally for context budgets at Opus 4.7. It is built around one assumption: maximal context is not optimal context. Curated context is.
The intuitive move when you get a 1M-token window is to load the entire repo and let the model figure it out. This produces worse diffs, for three reasons:
Pattern collisions. A large repo has multiple ways of doing similar things. Some are deprecated. Some are aspirational. Some are mid-migration. The agent loaded with all of them produces a diff that mixes patterns — picking up the deprecated one in some places and the new one in others. That is not a good diff. That is a debugging task.
Distraction load. Even at 1M tokens, the model's attention is finite. Filling the context with irrelevant files dilutes its focus on the relevant ones. The agent spends compute deciding what to ignore instead of what to do.
Stale signal. Old code in the context teaches the agent old patterns. If your repo has six months of accumulated drift, loading all of it tells the agent the drift is normal. It is not. It is debt.
The 1M-token window is a tool. Maximal context is the wrong tool for almost every task.
A working context for an Opus 4.7 vibe coding task usually has six parts:
1. The CLAUDE.md scaffold. Project constraints, patterns, dependencies, review standard. This is the spine. Every task starts here.
2. The spec. The thing the agent is being asked to do. Concrete acceptance criteria, negative criteria, test plan.
3. The files in scope. The specific files that will change. Loaded in full.
4. Interface surfaces. The public types, function signatures, and contracts the in-scope files depend on. Loaded as headers, not full implementations.
5. The relevant subset of recent PRs. If there is institutional context — a refactor in progress, a deprecation that is half-rolled — load it. Otherwise skip.
6. The relevant subset of tests. Tests that exercise the changing surfaces. Not the entire test suite.
That is usually 50K to 200K tokens for a substantial change — well under the 1M cap and well-curated for the work.
There are tasks where loading more context genuinely helps:
Codebase-scale audits. Security reviews, dependency audits, pattern surveys. These benefit from breadth because the question is breadth.
Cross-cutting refactors. When the change touches 30+ files in similar ways, loading them all helps the agent maintain consistency.
Investigating unfamiliar territory. When you do not know what the right files are, broader context helps the agent find them and explain back what it found.
For these tasks, fill the window. For everything else, do not.
The most expensive failure mode at Opus 4.7 is "context inflation." A team gets the bigger window, starts loading more context out of habit, and stops curating. Diff quality drops. The team blames the model. The model is fine. The discipline broke.
The discipline that prevents this is the same discipline that made smaller context windows work: ask, on every task, what does the agent need to see to do this well? The answer is rarely "everything." The answer is rarely the same as last task. Curating is the work.
At BridgeMind, context curation is treated as infrastructure. The CLAUDE.md scaffold is in the repo, versioned, reviewed when it changes. Spec templates are in the repo. The list of "interface surface" files for major modules is in the repo.
That sounds heavy. In practice it is the lightest part of the operating model. The alternative — re-curating context from scratch on every task, with every engineer making different choices — is what is heavy. It just looks light because the cost is hidden.
The Vibecademy certifications treat context engineering as one of the foundational competencies. The Claude Code track in particular drills it. Engineers earn the credential by demonstrating context budgets in reviewed work — not by reciting them.
Most of this strategy ports to other models. Curated context beats maximal context everywhere — at GPT-5.5, at Gemini, at any frontier model. The numbers change. The discipline does not. The multi-model vibe coding piece covers the routing logic for teams running more than one.
Opus 4.7's 1M-token window is special in the sense that it gives you more headroom than you usually need. It is not special in the sense that it changes what good context looks like. Good context still looks like good context.
If you have moved your team to Opus 4.7 and noticed diff quality dropping, the first thing to check is your context curation, not the model. The most common cause of post-upgrade quality regression is overcorrecting toward maximal context.
Pull the curated-context pattern in this piece into one of your repos. Run a week of work under it. Compare diff quality. The signal will be obvious.
The window is a tool. The strategy is the work. Vibecademy credentials the strategy. The window came free.
Continue Reading
Vibe coding is the practice of building software by describing intent to AI agents instead of writing every line by hand. Here is what that means for engineers shipping production code.
Agentic coding moves AI from suggestion engine to autonomous executor. Learn how agentic workflows differ from traditional AI assistance and what engineers need to know.
A practical playbook for vibe coding with the three tools that define agentic development. Workflows, patterns, and when to use each tool.