Context Efficiency
Efficiency removes waste, never evidence. It must not reduce necessary reasoning, security review, or verification.
Load by decision
- Start with repository instructions, the target symbol, its callers, and its focused tests.
- Search filenames, symbols, call sites, and Git history before opening files; read narrow ranges around matches rather than entire files.
- Load one capability or reference when it changes the next decision. Do not preload the catalog.
- Keep dependencies, lockfile bodies, generated output, build artifacts, and large logs out of context unless directly required. Store large raw outputs as files and inspect slices or summaries.
Avoid repeated work
- Do not reread unchanged files or rerun unchanged passing checks.
- Avoid duplicate agents, duplicate reviews, and overlapping skills.
- Prefer deterministic local scripts for repeatable parsing, discovery, counting, and validation.
- Run targeted tests first; broaden only when change risk requires it.
Spend context where risk lives
Keep API contracts, security boundaries, migrations, concurrency, and failures visible. Compress boilerplate and settled facts, not unresolved evidence.
Stop conditions
Stop exploring when acceptance criteria, affected boundaries, and a falsifiable implementation path are clear. Re-open context when a test contradicts the model, a hidden caller appears, or scope changes.
Character-based token estimates are useful for regressions but are not provider billing measurements or quality evidence. Delegation (delegation-discipline) is justified only when isolation or parallelism saves more than coordination costs.
1---2name: context-efficiency3description: >-4 Tactics for large repositories, noisy tool output, long sessions, and token5 pressure: search before reading, read ranges not files, keep logs and6 generated output out of context, never reread unchanged files or rerun7 unchanged passing checks, prefer deterministic scripts. Use when context8 volume materially affects execution quality; never reduces necessary9 reasoning, security, or verification.10license: MIT11metadata:12 source: https://skillmd.com/skills/al-janef/context-efficiency13---1415# Context Efficiency1617Efficiency removes waste, never evidence. It must not reduce necessary reasoning, security review, or verification.1819## Load by decision2021- Start with repository instructions, the target symbol, its callers, and its focused tests.22- Search filenames, symbols, call sites, and Git history before opening files; read narrow ranges around matches rather than entire files.23- Load one capability or reference when it changes the next decision. Do not preload the catalog.24- Keep dependencies, lockfile bodies, generated output, build artifacts, and large logs out of context unless directly required. Store large raw outputs as files and inspect slices or summaries.2526## Avoid repeated work2728- Do not reread unchanged files or rerun unchanged passing checks.29- Avoid duplicate agents, duplicate reviews, and overlapping skills.30- Prefer deterministic local scripts for repeatable parsing, discovery, counting, and validation.31- Run targeted tests first; broaden only when change risk requires it.3233## Spend context where risk lives3435Keep API contracts, security boundaries, migrations, concurrency, and failures visible. Compress boilerplate and settled facts, not unresolved evidence.3637## Stop conditions3839Stop exploring when acceptance criteria, affected boundaries, and a falsifiable implementation path are clear. Re-open context when a test contradicts the model, a hidden caller appears, or scope changes.4041Character-based token estimates are useful for regressions but are not provider billing measurements or quality evidence. Delegation (`delegation-discipline`) is justified only when isolation or parallelism saves more than coordination costs.