diff --git a/skills-manifest.json b/skills-manifest.json
index b4e8c78d8a..2f6ec2c094 100644
--- a/skills-manifest.json
+++ b/skills-manifest.json
@@ -26,7 +26,7 @@
       "files": 99
     },
     "hyperframes-cli": {
-      "hash": "f1c8c70693101e59",
+      "hash": "797fd624f0b83cb5",
       "files": 7
     },
     "hyperframes-core": {
diff --git a/skills/hyperframes-cli/SKILL.md b/skills/hyperframes-cli/SKILL.md
index 6e8d73fabe..6dfa2ddf8b 100644
--- a/skills/hyperframes-cli/SKILL.md
+++ b/skills/hyperframes-cli/SKILL.md
@@ -39,7 +39,14 @@ Cross-cutting rules that hold for every command:
 - **Use Studio context for user-directed edits.** When the user says "this selected element", "the thing I clicked", "current selection", or similar, ask them to select it in Studio, then run `npx hyperframes preview --context --json --context-fields selection`. Use the returned `selection.target.hfId` / `selector`, `selection.sourceFile`, `selection.currentTime`, and `selection.thumbnailUrl` to anchor the edit. If `selection` is `null` and `errors.selection.code` is `no-selection`, ask the user to click the element and rerun; do not guess from screenshots.
 - **Keep Studio context compact.** `preview --context --json` returns compact selection by default. Add `--context-fields selection`, `--context-fields selection,lint`, or `--context-fields lint` to avoid bloating agent context. Use `--context-detail full` only when you need heavy fields like computed styles, inline styles, or text-field metadata.
 - **Post-render verification.** After `render` returns exit 0, confirm the output file exists and has plausible size before reporting success: `[ -s "$OUTPUT" ] || echo "render produced no output"`. The CLI prints `◇  <path>` on success; for long renders also sanity-check duration with `ffprobe -i "$OUTPUT" -show_format -v error`.
-- **Report feedback after a successful render.** Once the output is verified, run `npx hyperframes feedback --rating <1-5> --comment "..."` once per task to report how the run went — `5` clean, lower when you hit friction. Put any bug, workaround, missing feature, or confusing behaviour you encountered in `--comment` (include the failing composition pattern and what you tried). This is the project's primary signal channel; renders that go silent leave the maintainers blind. Skip only when telemetry is disabled or the user opted out.
+- **Report feedback after a successful render.** Once the output is verified, run `npx hyperframes feedback --rating <1-5> --comment "..."` once per task to report how the run went — `5` clean, lower when you hit friction. Put any bug, workaround, missing feature, or confusing behaviour you encountered in `--comment`. This is the project's primary signal channel; renders that go silent leave the maintainers blind. Skip only when telemetry is disabled or the user opted out.
+  - **For a reproducible bug, add `--file-issue`** — it publishes a minimal repro of the project to a public URL and opens a pre-filled GitHub issue with your comment + environment. A comment alone almost never lets a maintainer reproduce; the published composition is what makes the report actionable. Point `--dir <project>` at the failing composition if it isn't the current directory.
+  - **Write bug comments a maintainer can reproduce from — a paraphrase they can't act on wastes the report.** The CLI already attaches os/arch/mem/ffmpeg/CLI-version, so don't repeat those; spend the comment on what only you saw. For a bug, include every one of these that applies:
+    - **Exact error string, verbatim** (copy it, don't summarise) — and whether the render still produced an output file, silently fell back (look for a `[Render] … re-rendering via screenshot` / `falling back` log line), or the process hard-exited with no error (`Killed: 9`, `FATAL … heap`, `Target closed`). "Which timeout / which failure mode" is the single most useful fact.
+    - **A minimal repro**: the smallest composition (or just the offending element + its CSS) that still shows it — inline it in the comment or note the file. Name the trigger you isolated (e.g. "only when the text node uses `font-family: 苹方`", "only at 4K", "only with a `<video>` present").
+    - **The exact command + flags** you ran, and any env vars set (`HF_*` / `PRODUCER_*`).
+    - **What's wrong and where**: the frame number or timestamp, and the visual defect (blank text, black band at the bottom, wrong glyphs, frozen canvas) vs. what you expected. A screenshot path if you saved one.
+    - If it reproduces only sometimes, say so and give the hit rate.
 
 ## Routing
 
diff --git a/skills/hyperframes-cli/references/preview-render.md b/skills/hyperframes-cli/references/preview-render.md
index 212910fcef..611a914d03 100644
--- a/skills/hyperframes-cli/references/preview-render.md
+++ b/skills/hyperframes-cli/references/preview-render.md
@@ -149,7 +149,7 @@ npx hyperframes feedback --rating 5                              # clean run, no
 npx hyperframes feedback --rating 3 --comment "bg <video> renders grey in multi-scene; worked around with --format png-sequence"
 ```
 
-`--rating` is 1-5 (required); `--comment` is free text — use it for any bug, workaround, missing feature, or confusing behaviour, plus the composition pattern that triggered it and what you tried. Feedback is anonymous and attaches a `doctorSummary` (OS/Node/CPU/mem/ffmpeg) automatically. No-ops when telemetry is disabled.
+`--rating` is 1-5 (required); `--comment` is free text — use it for any bug, workaround, missing feature, or confusing behaviour. Feedback is anonymous and attaches a `doctorSummary` (OS/Node/CPU/mem/ffmpeg) automatically, so don't repeat env in the comment. For a **bug**, a paraphrase can't be reproduced — give the maintainer: the **exact error string verbatim** and whether the render still produced output / silently fell back (`re-rendering via screenshot`) / hard-exited (`Killed: 9`, `Target closed`); the **isolated trigger** ("only at 4K", "only with a `<video>`", "only when the text uses `font-family: 苹方`"); the **exact command + `HF_*`/`PRODUCER_*` env**; and the **frame/timestamp + visual defect** vs. expected.
 
 Hit a reproducible bug? Add `--file-issue` (optionally `--dir <project>` and `--yes` for non-interactive shells) to also publish a minimal repro to a public URL and open a pre-filled GitHub `bug` issue draft for a maintainer to file. This publishes the project publicly, so it is opt-in and consent-gated; the issue is never auto-submitted.