Context window management
Skill charlieviettq/awesome-agent-skill/.claude/skills/context-window-management
Curated skill pack for LLM agents in engineer and science workflow (Cursor & Claude ready).
npx -y skills add charlieviettq/awesome-agent-skill --skill context-window-managementAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 22 stars22 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Manage LLM context budgets—prioritization, summarization, compaction, and what to load vs reference. Use for long sessions, large repos, or multi-doc tasks.
SKILL.md
2.3 KB, as published. Nobody here has run it
Context window management
Lifecycle
| Phase | Action |
|---|---|
| Bootstrap | Load spec, constraints, key paths only—not whole repo |
| During task | Append deltas; replace full file reads with summaries after use |
| Before compaction | Write Session state block (objective, done, open, paths) |
| Session end | Drop tool noise; keep decisions and verification evidence |
Strategy order
- Do not load — use search/read targeted files only.
- Reference — pointers (paths, line ranges) instead of full paste.
- Summarize — compress completed work; keep decisions and open items.
- Drop — remove obsolete tool output and duplicated content.
What to keep in active context
- Current objective and acceptance criteria
- Constraints (env, versions, "do not touch")
- Recent errors and fixes
- Open questions
What to offload
- Full file contents already summarized
- Long command outputs (keep exit code + last N lines)
- Exploratory branches that did not ship
Compaction pattern
## Session state (compacted)
Objective: ...
Done: ...
Open: ...
Key paths: ...
Do not repeat: [list already-established facts]
Multi-file work
- Load one module at a time when possible.
- Prefer ripgrep/semantic search over dumping trees.
- For RAG: retrieve only top-k relevant chunks.
Anti-patterns
- Pasting entire repos or logs into chat.
- Re-explaining architecture every turn without delta.
- Keeping failed attempts at full length after resolution.
Session setup (context engineering)
Before long tasks, establish a minimal working set:
- Objective block — one paragraph: goal, non-goals, done-when
- Constraint block — versions, env, files off-limits
- Reference map — paths to spec, ADR, or ticket; load on demand
- Tool policy — which tools/skills apply; avoid loading unused skill bodies
Refresh the objective block after major pivots; do not re-paste full history.
Pairing
Use with rag-systems for external knowledge; clarify-underspecified when scope grows mid-session; interview-me before loading large specs.