# Ackit Context Optimization > Build budgeted context packs and trim instruction bloat for coding agents. - Skill: `cynrath/ackit-context-optimization` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add cynrath/ackit-context-optimization` - Raw SKILL.md: https://api.skillmd.com/api/skills/cynrath/ackit-context-optimization/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: Cynrath (https://skillmd.com/u/cynrath) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/cynrath/ackit-context-optimization --- # Context Optimization Activate when the user asks to prepare context for an agent session or reduce token usage. See [ranking signals](references/ranking.md) for deterministic ordering. ## Steps 1. `ackit pack --max-tokens [--format markdown|json] [--include ] [--changed] [--profile codex|claude|copilot|gemini|generic] [--task ] [--resume]` to generate a deterministic pack with a manifest of included/excluded files and reasons (`relativePath`, `action`, `reason`, `estimatedTokens`, `sha256`, `bytes`). Task-aware ranking (`--task`) and checkpoint resume (`--resume`) keep packs scoped; greedy fill excludes over-budget candidates with `budget exhausted`. 2. Review exclusions; add explicit includes only when the ranking missed real relevance (explicit include is the highest signal; no embeddings). 3. For bloated instructions, run `ackit optimize [--profile ] [--category ] [--min-severity low|medium|high] [--explain] [--format terminal|json|markdown|sarif]` (read-only by default) and apply suggestions manually or with `ackit optimize --fix [--dry-run] [--diff]` — fixes apply ONLY to ACKit-managed surfaces. ## Notes - Token counts are character-class estimates (~4 chars/token); treat budgets as soft targets. - Safety gates run before scoring: secret-shaped content excluded, duplicate hashes deduped, machine-local paths scrubbed to ``. - Never paste repository content into external services.