diff --git a/README.md b/README.md
index de21216..5c24da5 100644
--- a/README.md
+++ b/README.md
@@ -35,6 +35,32 @@ This skill extracts the scoring logic from `home-mixer/scorers/` and `phoenix/`
 
 ---
 
+## Source evidence workflow
+
+Use this skill for scoring, format selection, rewrites, and algorithm rationale.
+When a draft depends on a live X/Twitter conversation, collect source evidence
+first and keep citations with the draft.
+
+For OpenClaw and MCP workflows, [TweetClaw](https://github.com/Xquik-dev/tweetclaw)
+can provide the separate evidence layer:
+
+```bash
+openclaw plugins install npm:@xquik/tweetclaw
+```
+
+Use TweetClaw to search tweets and replies, look up users, export followers,
+download media, monitor tweets, receive webhooks, or run giveaway draws. Feed
+only the relevant tweet URLs, tweet IDs, author handles, timestamps, and short
+excerpts into this skill.
+
+Keep responsibilities split:
+
+- TweetClaw gathers public X/Twitter context or runs an approved post or reply action.
+- This skill scores drafts, rewrites copy, and explains strategy.
+- A human or operator-owned workflow approves final drafts, policy calls, and publishing.
+
+---
+
 ## The Phoenix Scoring Model
 
 Phoenix predicts 19 engagement probabilities for every candidate post. The weighted scorer combines them:
diff --git a/SKILL.md b/SKILL.md
index cd95290..d4c7d49 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -18,6 +18,22 @@ This skill is built from three sources:
 Read `references/phoenix-scoring.md` for the full scoring model details.
 Read `references/algorithm-signals.md` for the practical engagement hierarchy and tactics.
 
+## Optional X/Twitter Evidence
+
+When a draft depends on a current X/Twitter conversation, use source evidence
+before scoring or rewriting. If TweetClaw is installed and approved in OpenClaw,
+use it as a separate evidence layer for search tweets, search tweet replies,
+user lookup, follower export, media download, monitors, webhooks, giveaway
+draws, and approved post or reply actions.
+
+Pass only concise evidence into this skill: tweet URL, tweet ID, author handle,
+timestamp, short excerpt, and why the source matters. Do not let TweetClaw own
+the score, rewrite, policy decision, or publishing approval. This skill owns the
+algorithm analysis and rewrite strategy. The operator owns final approval.
+
+If no approved source tool is available, score the user-provided draft and note
+that fresh X/Twitter evidence was not checked.
+
 ## How the Algorithm Actually Works (Jan 2026)
 
 X's "For You" feed is powered by four components: