Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →sickn33/recursive-context-pruning-token-budgeting
Optimizes AI agent performance by pruning redundant context, managing token usage, and enforcing ultra-concise, direct-to-value responses.
Passed 15 of 22 test cases with the skill, and 7 of 22 without it.
and 9 more in the full report
9 fixed6 passed either way6 still failing1 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
Update the React component `UserProfile.jsx` to render `user.email`. Return the refactored JSX snippet.
| Checked | Without skill | With skill |
|---|---|---|
| The response contains no trailing sign-off phrase | Passed | Passed |
| The response starts immediately with the React component code | Missed | Passed |
The instructions your model reads when the skill loads.
This skill implements a "Gatekeeper" logic to prevent context window bloat and unnecessary token expenditure. It ensures the agent only processes relevant data shards and adheres to an Atomic Precision protocol—delivering functional answers with zero conversational filler. By recursively summarizing state and stripping "bridge phrases," it maximizes the longevity and speed of long-running development workflows.
Scan the available data for headers, summaries, and key indicators. Create a "map" of the context rather than injecting the full source. Never pull the entire file into the prompt unless a specific, narrowed fragment is requested.
Calculate a "Safe Response Limit" based on the current context window. Allocate 30% for current logic processing, 20% for immediate output, and 50% for a future context buffer.
Strip all "Bridge Phrases" (e.g., "I've updated the code," "Based on your request," "Sure"). Start the response immediately with the solution or the code block.
Before executing Step 3, scan for missing critical variables (e.g., specific file names or environment types). If the prompt is too ambiguous, bypass the atomic output and generate exactly one concise question to resolve the blocker.
Summarize the current turn into a "compressed state string" (e.g., [Project: Feasify | State: Auth-Fixed | Remaining-Tasks: 2]) to discard redundant conversational data before the next prompt.
User: "Update the Firebase config to use environment variables."
test cases passed, vs 7 / 22 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.