Prevent Redundancy

skill: prevent-redundancy

purpose

Before sending a response, remove content that was already said earlier in the conversation. Eliminates the agent habit of re-explaining things the user already knows, restating conclusions, and padding with information already on screen.

whentouse

• Multi-turn conversations where context has accumulated
• Agent loop where the same output type is generated repeatedly
• Before summarizing or concluding — summaries often just repeat the body

whennotto_use

• First turn with no prior context
• User explicitly asks you to repeat or restate something
• Generating formal documents where repetition serves structure (exec summary, conclusion)

input

draft: string            # What you're about to say
prior_context: string    # What has already been said this session

output

{
  "redundant_parts": [
    { "text": "string", "already_said_in": "string" }
  ],
  "clean_draft": "string",
  "reduction_pct": "int"
}

instructions

For each claim or statement in the draft: has this already been established in prior_context?

Redundant if: restates a fact already established, re-explains something the user confirmed understanding of, summarizes points just made in the same response.

NOT redundant if: adds a new angle or detail, briefly mentioned before but now developed meaningfully, direct answer to a new question even if related to prior topics.

Remove redundant parts. Do not replace them — just cut.
clean_draft must still be a complete, coherent response.
If nothing is redundant, redundant_parts is empty and clean_draft equals the input draft.
If the entire draft is redundant: clean_draft should be a single sentence that acknowledges and moves forward (e.g. "Got it — moving on." or "Confirmed. Proceeding with X.").

Return JSON only.

constraints

• Max output tokens: 400
• clean_draft must be coherent after cuts
• Only flag genuine repetition — not related content or elaboration

example

Input:

draft: "To summarize: the issue is a memory leak from unclosed DB connections. As I mentioned, the connection pool isn't released after each request. The fix is adding connection closing in the finally block, which I described. This resolves the memory leak."
prior_context: "Agent turn 3: Memory leak caused by unclosed DB connections — pool not released after requests. Fix: connection.close() in a finally block."

Output:

{
  "redundant_parts": [
    { "text": "the issue is a memory leak from unclosed DB connections", "already_said_in": "Turn 3 established root cause" },
    { "text": "the connection pool isn't released after each request", "already_said_in": "Turn 3 explained mechanism" },
    { "text": "adding connection closing in the finally block, which I described", "already_said_in": "Turn 3 gave the fix" },
    { "text": "This resolves the memory leak", "already_said_in": "Restates conclusion already made" }
  ],
  "clean_draft": "Fix is in — connection.close() in the finally block. Ready to test.",
  "reduction_pct": 76
}

feedback_log

2026-05-29 | entire draft repeated prior turn verbatim | constraint said clean_draft must be coherent but nothing was left after cuts — undefined behavior | added: when 100% redundant, output one forward-moving sentence

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: 0-uddeshya-0
• Source: 0-uddeshya-0/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.