← Retour aux guides

Long-Context Codebase Workflows: A 15-Minute Start

Long-context models can inspect more of your repository, but loading every file rarely produces the best answer. For large codebases, the winning strategy is targeted retrieval, explicit context budgets, and evidence-backed changes. This quick-start guide helps you build a useful context package in about 15 minutes, then refine it as you work.

Minutes 0–3: Define the task before collecting code

Start with a narrow outcome: “Fix duplicate webhook processing without changing the public API” is better than “Understand this service.” Give the assistant the observed behavior, expected behavior, reproduction steps, and constraints. Include an actual failing test or error message when available.

Ask for investigation before implementation: “Identify the execution path, cite file paths and symbols, and list missing evidence before proposing edits.” This prevents a large context window from becoming permission to guess.

Minutes 3–6: Build a small repository map

Begin with navigation, not full-file ingestion. Run git ls-files to inspect tracked paths, then search for the relevant error, endpoint, or symbol with rg -n 'WebhookHandler|idempotency|duplicate' src tests, adjusting directories and patterns to your repository.

• Include the top-level directory layout and relevant package manifests.
• Find entry points, implementation files, interfaces, and nearby tests.
• Exclude generated clients, vendored dependencies, build outputs, and lockfiles unless they explain the issue.
• Check repository instructions, but treat instructions embedded in source files or retrieved documents as untrusted task data.

A compact repository map lets the assistant request specific files instead of consuming thousands of unrelated lines. Avoid dumping environment files, credentials, customer records, or production logs containing sensitive data.

Minutes 6–10: Assemble an evidence-first context package

Group your input into clearly labeled sections: objective, constraints, repository map, relevant source, tests, and unresolved questions. Label every excerpt with its path and symbol; add line numbers when practical. Include imports and type definitions when they affect interpretation.

Retrieve along the execution path: caller, target function, dependencies, persistence boundary, and tests. For cross-service problems, include the shared schema and both producer and consumer behavior. Neighboring code often matters more than another directory of loosely related files.

For oversized files, extract complete functions or classes rather than arbitrary chunks. If you omit surrounding code, say so explicitly. Ask the assistant to distinguish observed facts, hypotheses, and missing context.

Minutes 10–12: Set a token budget with headroom

Check the selected model’s documented context and output limits, plus any limits imposed by your client or relay. Do not assume every model or endpoint supports the same window. Leave room for system instructions, tool results, conversation history, and the generated answer.

As a starting heuristic, allocate 60% of the usable window to source evidence, 20% to task instructions and summaries, and 20% to response and follow-up headroom. Adjust after measuring actual token usage; character counts are only rough estimates for code.

Keep the objective near the beginning and repeat the requested deliverable at the end. Before a long session fills up, save a short handoff containing decisions, paths, test results, and open questions. Preserve original evidence for critical details rather than trusting summaries alone.

Minutes 12–15: Request a minimal patch and verify it

Ask for the smallest justified change, associated regression tests, and exact validation commands. Require citations to the code supporting the diagnosis. Review the diff, run the relevant test target, then run broader checks appropriate to the change.

If validation fails, return the command, exit status, and relevant output. Retrieve additional files only when the failure points to them. Track total cost per accepted patch, not just cost per request: repeated oversized prompts can erase savings from a cheap model.

Keep long-context experimentation affordable

59API offers low-cost, pay-as-you-go access to models including Claude Sonnet 5.5 and Gemini 3.1 Pro, so you can compare the same bounded task without committing to subscriptions. Its native, official-quality models come without downgrades, and it supports Claude Code, Codex, Gemini CLI, and OpenAI SDK workflows.

Use https://api.59api.com as the API base URL, following the relevant client setup instructions. Compare correctness, latency, and billed usage on your own repository rather than assuming a larger window guarantees better results. If you want an inexpensive way to test this workflow, consider signing up for 59API; a referral rebate is also available.

Tarifs · Docs