Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Is that skill safe to install?
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
Are you the author of Efficient Context?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/rabee-elkholy-efficient-context)
---
name: efficient-context
description: Bounded graph-backed context resolution (`task-context --since`, batch targets) and Smart Context Bridge (`task bridge --save`) for zero-bloat Antigravity sessions.
version: 1.1.0
kernel-major: 1
---
# Efficient Context & Smart Bridge Skill
Follow `.agents/workflows/efficient-context.md` and `.agents/rules/pm-and-context-bridge.md`. Batch up to 8 targets in one `task-context` call, pass `--since <context_digest>` only within confirmed live turns, and execute `task bridge --save` / `task bridge` on the developer's behalf without printing raw CLI bridge commands in chat.
## Hybrid Quota-Safe Delegation & Anti-Reread Contract (Option 4)
- Keep the primary chat on `Gemini 4 Carbon (Medium)` and rotate to `Claude Sonnet 5.5 / Opus 5.5` when the `/usage` Gemini weekly bar reaches ~50%.
- **Parallel Flash Scouts (`harness-scout-agent`)**: Delegate multi-file codebase discovery (pre-plan or mid-task) to `harness-scout-agent` (`gemini-3.8-flash-medium`; never built-in `research` or `explore`). When investigating `3+` distinct subsystems or screens, split into `2–3` parallel `harness-scout-agent` entries in a single `invoke_subagent` call (`<= 24` tool calls per scout, `4–8` parallel `view_file` reads per turn).
- **Scoped Flash Worker (`harness-worker-agent`)**: During `IMPLEMENTING`, check `python .agents/harness.py delegate-plan --json` and dispatch a single `harness-worker-agent` (`Workspace='inherit'`) with a Decision-Complete Brief for `T0-T3` non-sensitive slices and Gradle compile/test loops; retain `T4/T5` sensitive production edits (`AUTH`, `BILLING`, `SECURITY`, `CRYPTO`) on the Lead Architect.
- **Zero Post-Subagent Re-Read Across All Waves**: Once ANY `harness-scout-agent`, `harness-worker-agent`, specialist, or reviewer wave returns, synthesize its file:line digest directly (or run `git diff` once after a worker slice, or `review complete-all` / `phase-review complete-all` after reviewers) without re-running `graph` or re-reading analyzed source files (hard cap: `<= 3` surgical `<= 200`-line gap checks per wave).
- **Surgical Symbol Tools on Large Files (`>350` lines)**: Use `python .agents/harness.py symbol outline --file <path>` and `symbol read --file <path> --symbol <Class.method>` (or `symbol replace-body` / `symbol insert-after` with `--expected-sha256`) to inspect/edit single declarations without loading 800+ surrounding lines.
- When the developer asks why a chat stalled or burned tokens, run `python .agents/harness.py diagnose-session --json` (add `--export-bundle` only if they request a Zero-PII diagnostic file).