Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Preview a distilled wiki page from inside Claude Code. Prints title, summary, source count, and the local md path. Full body lives on disk — open with the user's editor. Invoked as `/read <title_or_id>`.
.claude/skills/davepoon-read/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 127% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 78% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 78% | 0% |
Surface a page's identity so the user can decide whether to open it. /read is preview, not full text. The body is on disk; preview keeps chat scannable, dodges Bash output truncation, and respects the "md is canonical, viewer is the user's editor" model.
Two shapes accepted:
page_; legacy concept_ ids still work) → direct fetch.Both end with the same preview block.
Call the MCP tool to fetch the page:
get_page(page_id="<id>")The response is a JSON object wrapping { "page": {...} }. Read title, summary, space, and source_memory_ids off the page, then look up the md filename in ~/.origin/pages/.origin/state.json:
Bash: python3 -c '
import json, os, sys
state_path = os.path.expanduser("~/.origin/pages/.origin/state.json")
pid = "<id>"
filename = None
try:
with open(state_path) as f:
filename = json.load(f).get("pages", {}).get(pid, {}).get("file")
except FileNotFoundError:
pass
print(f"~/.origin/pages/{filename}" if filename else "(no md projection on disk)")
'Combine the page fields with the resolved md path and emit the preview block.
Search first, then fetch the top hit by id and run the same preview block. Always resolve through the search_pages MCP tool — never derive a slug client-side (skill heuristics drift from the canonical slugify() on apostrophes and punctuation).
search_pages(query="<arg>", limit=1)Parse the returned JSON — the response shape is { "pages": [...] }. Read the first hit's id. If pages is empty, tell the user "no page found matching <arg> — try /distill <arg> to create one" and stop. Otherwise call get_page(page_id=<id>) and emit the same preview block from section 1.
Always print exactly these lines (no body):
Title: <title>
Version: v<N> — <last_edited_by> <relative_time> (<last_delta_summary>)
Summary: <one sentence>
Sources: <N> memories
Space: <space or (none)>
Links: <N inbound, M outbound (<K> broken)>
Open: ~/.origin/pages/<slug>.md
⚠ Stale: <stale_reason> — run /distill to refreshLock banner rule:
When the page payload returned by get_page has user_edited == true, prepend this line as the very first line of the rendered output, before the title:
🔒 You've edited this page. Auto-refresh paused. `/distill rebuild <page-id>` to unlock.Substitute <page-id> with the actual page.id value. When user_edited is false or absent, omit this line entirely.
The lock means daemon distill cycles will not auto-rewrite this page's prose from sources — edits stay until the user explicitly runs /distill rebuild to wipe and regenerate.
Version line rules:
v<N>. When version is null or missing, omit the line. — <last_edited_by> when populated (e.g. re_distill, user, agent).last_edited_at is set (e.g. 2h ago, 3d ago).(<last_delta_summary>) when the field is non-empty.v1 — synthesized 4h agov4 — re_distill 2h ago (+mem_xyz, +250 chars)v3 — user 1d agoStale warning rule:
⚠ Stale: line only when stale_reason is non-null/non-empty.stale_reason verbatim (daemon sets it; values likesource_updated, new_memories are human-readable enough).
Links line rules (unchanged): Call get_page_links(page_id="<id>") right after get_page and count:
len(inbound)len(outbound)target_page_id is nullOmit the parenthetical when broken is zero. Drop the line entirely if inbound + outbound is zero.
Don't paraphrase the title, don't trim the summary, don't decorate the preview. The block is one screen, predictable, easy to skim.
If the user wants the full body, they open the md file in their editor (Obsidian, VS Code, glow, bat, …). The plugin doesn't render markdown better than their tools do.
/distill finishes, follow up with /read "<title>" (titlessurvive any rendering surface; ids may visually truncate).
/recall.list_pages_recent MCP tool orls ~/.origin/pages/.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 6,003 | 4,157 | -31% | 1 | 1 | 0% | 1,127 | 1,630 | +45% | 0 | 0 | — |
case-02 | fail→fail | 4,597 | 4,963 | +8% | 1 | 1 | 0% | 942 | 1,638 | +74% | 0 | 0 | — |
case-03 | fail→fail | 7,041 | 4,414 | -37% | 1 | 1 | 0% | 590 | 1,585 | +169% | 0 | 0 | — |
case-04 | fail→pass | 5,396 | 6,697 | +24% | 1 | 1 | 0% | 816 | 1,853 | +127% | 0 | 0 | — |
case-05 | fail→fail | 3,776 | 4,671 | +24% | 1 | 1 | 0% | 620 | 1,630 | +163% | 0 | 0 | — |
case-06 | fail→fail | 8,966 | 1,764 | -80% | 1 | 1 | 0% | 1,389 | 1,672 | +20% | 0 | 0 | — |
case-07 | pass→pass | 10,923 | 1,838 | -83% | 1 | 1 | 0% | 1,808 | 1,679 | -7% | 0 | 0 | — |
case-17 | pass→pass | 5,585 | 2,535 | -55% | 1 | 1 | 0% | 933 | 1,715 | +84% | 0 | 0 | — |
case-08 | fail→pass | 8,282 | 986 | -88% | 1 | 1 | 0% | 1,512 | 1,513 | +0% | 0 | 0 | — |
case-09 | pass→pass | 9,181 | 2,247 | -76% | 1 | 1 | 0% | 1,737 | 1,824 | +5% | 0 | 0 | — |
case-10 | pass→pass | 8,748 | 1,887 | -78% | 1 | 1 | 0% | 1,466 | 1,704 | +16% | 0 | 0 | — |
case-11 | pass→pass | 8,914 | 1,538 | -83% | 1 | 1 | 0% | 1,484 | 1,607 | +8% | 0 | 0 | — |
case-12 | pass→pass | 8,125 | 2,426 | -70% | 1 | 1 | 0% | 1,407 | 1,671 | +19% | 0 | 0 | — |
case-13 | fail→pass | 5,673 | 1,442 | -75% | 1 | 1 | 0% | 916 | 1,635 | +78% | 0 | 0 | — |
case-14 | pass→pass | 2,681 | 1,990 | -26% | 1 | 1 | 0% | 446 | 1,652 | +270% | 0 | 0 | — |
case-15 | pass→pass | 3,434 | 1,883 | -45% | 1 | 1 | 0% | 565 | 1,693 | +200% | 0 | 0 | — |
case-16 | fail→pass | 7,347 | 1,631 | -78% | 1 | 1 | 0% | 1,277 | 1,600 | +25% | 0 | 0 | — |
case-18 | fail→pass | 6,028 | 2,590 | -57% | 1 | 1 | 0% | 983 | 1,751 | +78% | 0 | 0 | — |
case-19 | fail→pass | 12,378 | 1,897 | -85% | 1 | 1 | 0% | 2,101 | 1,666 | -21% | 0 | 0 | — |
case-20 | pass→pass | 5,136 | 2,331 | -55% | 1 | 1 | 0% | 865 | 1,745 | +102% | 0 | 0 | — |
case-21 | fail→pass | 10,057 | 2,576 | -74% | 1 | 1 | 0% | 1,753 | 1,770 | +1% | 0 | 0 | — |
case-22 | fail→pass | 6,665 | 2,691 | -60% | 1 | 1 | 0% | 1,210 | 1,849 | +53% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.