diff --git a/README.md b/README.md
index 8534814..ce088d5 100644
--- a/README.md
+++ b/README.md
@@ -5,13 +5,6 @@ the answer on stdout, diagnostics on stderr, a verdict in the exit code. Speaks
 to any OpenAI-compatible endpoint. Sibling binary `clank-jev` routes and gates;
 compose them — do not fold chat/agent UX into `clank`.
 
-```sh
-cargo build --release
-export CLANK_BASE_URL=http://127.0.0.1:8080/v1   # your server — no laptop-only default
-export CLANK_MODEL=$(curl -s "$CLANK_BASE_URL/models" | jq -r '.data[0].id')
-./target/release/clank -m 'reply with exactly: pong'    # -> pong, exit 0
-```
-
 ![clank in a shell](docs/images/clank-demo.gif)
 
 Recorded against a live server by `scripts/record-demo.sh`: observers, pipe as
@@ -86,23 +79,18 @@ two variables plus the flags in [Configuration](#configuration) are the whole
 configuration: no config file, no state on disk. Container instead of a
 toolchain: [In a container](#in-a-container).
 
-## What it is for
 
-Ask a model about text you already have, and get back something your shell can
-act on:
+## 30s smoke
+
+After install, point at your OpenAI-compatible server and prove the pipe:
 
 ```sh
-rg -n -C3 "userData" src/ | clank --thinking off -m "what does this do?"
-git diff | clank --thinking off --max-tokens 400 -m "review this diff, one line per issue"
-rg -n "TODO" src/ | clank --each --thinking off -m "one line: actionable now, or not?"
-clank --json-schema @schema.json -m "extract the findings" | jq -er .
-clank --jsonl -m "summarize" | clank -m "what did you say?"
-clank --system @prompts/review-sh.md -c script.sh -m "review the script"
+export CLANK_BASE_URL=http://127.0.0.1:8080/v1
+export CLANK_MODEL=$(curl -s "$CLANK_BASE_URL/models" | jq -r '.data[0].id')
+clank -m 'reply with exactly: pong'    # -> pong, exit 0
 ```
 
-## Keeping it fast
-
-### Harness cost (dated)
+## Harness cost (dated)
 
 **2026-09-22** one-shot on the same Go model (`opencode-go/glm-5.3-flash`),
 tools off, prompt `→ pong`. This is **pipe vs agent CLI**, not a quality
@@ -118,6 +106,22 @@ benchmark — agent harnesses still pay TUI/runtime tax with `--no-tools`.
 Full method, caveats, and runs:
 [docs/history/clank-vs-agents-2026-09-22.md](docs/history/clank-vs-agents-2026-09-22.md).
 
+## What it is for
+
+Ask a model about text you already have, and get back something your shell can
+act on:
+
+```sh
+rg -n -C3 "userData" src/ | clank --thinking off -m "what does this do?"
+git diff | clank --thinking off --max-tokens 400 -m "review this diff, one line per issue"
+rg -n "TODO" src/ | clank --each --thinking off -m "one line: actionable now, or not?"
+clank --json-schema @schema.json -m "extract the findings" | jq -er .
+clank --jsonl -m "summarize" | clank -m "what did you say?"
+clank --system @prompts/review-sh.md -c script.sh -m "review the script"
+```
+
+## Keeping it fast
+
 The model is the constraint (clank's own share is 1 ms of startup), so the lever
 is the tokens you pay for: