ckdan19/LabNotebook-Summarizer / .claude/skills/daily-literature-post
daily-literature-post
VerifiedDaily Literature Post Skill
Trigger this skill when the user asks to run the daily literature connection post, do today's literature connections, run the daily lit post, or any similar phrasing that means "look at today's lab posts and connect their findings to the published literature."
What this skill does
Fetches the last few days of lab notebook posts, keeps only the ones that describe a real scientific finding, runs the literature-connector skill against each finding, and assembles the results into a single Markdown post published to WordPress. Future daily runs exclude this skill's own posts primarily by a title-prefix match (titles beginning Daily Literature Connections —); the post is also tagged with the auto-literature-connections category as a second signal, which now sticks (an admin created the category on the site on 2026-08-25; the publishing token can attach the existing category but still cannot create categories — see step 3). Every post it processes is recorded in digests/.literature-state.json so it is never processed twice.
This is a narrow daily counterpart to full-lab-digest: it does not summarize the notebooks, run subagents, or search the archive — it only connects today's genuine findings to external literature.
Live publishing is authorization-gated; draft is the safe default. This skill publishes live only when a durable authorization file is present; otherwise it falls back to a draft. Two independent rails protect this (both enforced in step 7):
- Authorization file. Live publishing requires
AUTHORIZATION.mdat the repository root containing a standalone line reading exactly> AUTOMATED PUBLISHING: AUTHORIZED(see step 7b for the precise match rule — a mention of the phrase elsewhere in the file's prose does not count). If the file is missing, unreadable, or lacks that exact line, the skill publishes a draft and says so in its report. There is no way to publish live without this file — deleting or editing out the marker line instantly reverts the skill to draft-only. - One-post-per-day cap. The skill publishes at most one post per day (live or draft). If the state file already records a publish for today, it refuses to publish again rather than posting a duplicate.
Never publish live by any other route (e.g., hand-passing --status publish outside these rails). The rails exist so the pipeline stays safe when it runs unattended after the original author has left the lab. See AUTHORIZATION.md for who approved live publishing and how to pause or disable it.
State file — digests/.literature-state.json
This skill keeps persistent state in digests/.literature-state.json (path relative to the repository root), mirroring the union/persist pattern of digests/.digest-state.json. It has this shape:
{
"last_run_date": "2026-08-10",
"connected_urls": [
"https://genefish.wordpress.com/2026/08/09/some-real-finding/"
],
"deferred_urls": [
{"url": "https://genefish.wordpress.com/2026/08/09/a-finding-with-no-hits/", "first_deferred": "2026-08-10"}
],
"publish_log": [
{"date": "2026-08-10", "status": "draft", "url": "https://genefish.wordpress.com/?p=123"}
]
}
connected_urls is the set of every post URL this skill has finished with — either it produced a literature section, or it was evaluated twice and yielded nothing both times. It is committed to the repo — never add it to .gitignore — so de-duplication works across machines and collaborators.
deferred_urls holds posts that carried a real finding but whose literature-connector search came back empty on their first evaluation. An empty search is not the same as an absent finding: a narrow topic against literature-connector's 12-month recency window can return zero for a result that is genuinely connectable, and indexing lags mean the same query can hit a day later. Each entry records the url and the first_deferred date. A deferred post gets exactly one more evaluation (step 2), after which it moves to connected_urls whatever the outcome — so no post is evaluated more than twice, ever.
publish_log records every post this skill has published, one entry per publish, each with the date (YYYY-MM-DD) it was published, the status it was published with ("draft" or "publish"), and the resulting post url. It drives the one-post-per-day cap (step 7): the number of entries whose date equals today is the count of posts already published today.
- If the file does not exist (first run): treat
connected_urlsas an empty set andpublish_loganddeferred_urlsas empty lists, exclude nothing on that basis, and create the file in the state-update step below. - If
publish_logis missing from an older state file: treat it as an empty list (no publishes recorded yet) and add it when you next write the file. - If
deferred_urlsis missing from an older state file: treat it as an empty list and add it when you next write the file. - Match on the canonical post URL only. Treat URLs that differ solely by a trailing slash as the same URL.
Steps
1. Fetch today's posts
Run:
python3 scripts/fetch_lab_posts.py --days 3
If the invoker specified a window, use that number instead of 3. A request like "run
the daily-literature-post skill over a window of the last 9 days" means --days 9. This
is how a catch-up run is performed after an outage (see below). Absent an explicit
window, always use 3.
Why 3 days and not 1. The window is deliberately wider than the daily cadence. When
this skill runs unattended, a scheduled run can be delayed or skipped entirely (GitHub
Actions cron is best-effort, and a machine running cron can be asleep or offline). With
a 1-day window, a single missed run drops that day's findings permanently — the posts
fall outside every subsequent window and are never connected.
A wider window costs nothing because step 2's connected_urls check already guarantees a
post is processed at most once, ever. Re-reading a post that was already handled yesterday
is a no-op, so the overlap simply lets a late run catch up on what a missed run skipped.
Do not narrow this back to --days 1; the de-duplication, not the window, is what
prevents repeats.
The window is what bounds that self-healing, so know its limit. A gap longer than
the window is not recovered by it — those posts fall outside every subsequent window and
are lost exactly as they would be with --days 1. This is not hypothetical: a CI
permission fault stalled every run from 2026-08-25 to 2026-09-01, an 8-day gap against a
3-day window, and the posts from 2026-08-25 to 2026-08-28 were never evaluated. Recovering
from a gap that long requires an explicit wider window, which is what the paragraph above
is for; the daily workflow exposes it as a days dispatch input.
Parse the JSON posts array. Each post has author, date, title, url, content, and categories.
If posts is empty, exit quietly. Produce no post, write no file, do not touch the state file, and print nothing beyond a brief note that there were no posts today. This is a silent no-op, not an error.
2. Exclude already-processed posts (state file)
Read digests/.literature-state.json (see above). Drop any fetched post whose canonical url is already in connected_urls. These are finished with and must not be processed again.
Do not drop posts listed in deferred_urls. These carried a real finding but returned no literature on their first pass, and are being given their one retry. Keep them in the candidate set and carry them through steps 3–5 as normal; note which ones they are, because step 8 treats their outcome differently. A deferred post is only reachable while it still falls inside the fetch window from step 1 — if the window has moved past it, it simply never comes back, and step 8 retires it.
3. Exclude automated pipeline output (this skill's and full-lab-digest's)
Drop any remaining post that is automated pipeline output — either this skill's own prior drafts, or the full-lab-digest skill's published digests. These are aggregated/meta posts, not primary lab findings; connecting them would feed the pipeline its own (or a sibling pipeline's) output. Exclude a post if any of the following match:
- This skill's own output (title): the post's
titlebegins withDaily Literature Connections —(case-insensitive; the em dash is what step 6's H1 template emits → WordPress title, e.g.Daily Literature Connections — 2026-08-18). This is the reliable check for this skill's own output — match on it first. - This skill's own output (category): the post's
categorieslist (returned byfetch_lab_posts.py) contains the exact nameauto-literature-connections. Since 2026-08-25 the category exists on the site, so posts published by this skill with--category auto-literature-connectionsnow come back from the WordPress API with the category attached — this check works as a real second signal. The title check above is still the primary guard; keep both. - Full Lab Digest output (title): the post's
titlebegins withFull Lab Digest —(case-insensitive; the em dash is what thefull-lab-digestskill emits as its H1 → WordPress title, e.g.Full Lab Digest — 2026-08-04 to 2026-08-10 (7 days)). Match on the title prefix, not on content — do not rely on inferring "this is a digest" from the body alone. - Full Lab Digest output (category): the post's
categorieslist containsfull-lab-digest, if present.full-lab-digestis normally published viawordpress-publisherwithout a--category, so it typically lands asUncategorizedand this check will usually not fire — the title check above is the reliable one. Keep this category check anyway as belt-and-suspenders in case a digest is ever tagged.
This is a belt-and-suspenders check alongside step 2's state-file check — apply both. A post is excluded if the state file (step 2) or any of the tests above flag it. Do not treat this as a substitute for step 4's real-finding filter; a digest whose title check somehow fails would still be caught there, but that content-based catch is not guaranteed reliable long-term, which is why the explicit title/category exclusion above exists.
4. Keep only posts with a real scientific finding
For each surviving post, read its content and decide whether it describes a real scientific finding — a concrete experimental result, a measurement, an observation, a molecular/physiological outcome, a data analysis conclusion, etc.
Exclude posts that are purely logistical: meeting notes, scheduling or goals updates, equipment or reagent orders, inventory, lab-management notes, protocol drafting with no result yet, and similar. When a post has no genuine finding, skip it — do not invent one.
If every post today is logistical (nothing survives this filter), treat the day as empty: behave exactly like step 1's empty case — produce nothing, write nothing, touch no state — and report that nothing qualified today.
5. Derive TOPIC + FINDING and run literature-connector per post
For each surviving post with a real finding:
- Derive a concise TOPIC — the biological/scientific subject (e.g.,
"PolyIC immune priming in Pacific oysters"). - Derive a FINDING — one or two sentences describing the specific lab result from the post body (e.g.,
"VIPERIN is upregulated by PolyIC but HSP70 is not, suggesting pathway specificity.").
Then invoke the literature-connector skill exactly as if the user had typed:
/literature-connectorTOPIC: [topic phrase] FINDING: [finding description]
Run these searches sequentially, not in parallel, to avoid hitting the PubMed / Europe PMC rate limits. All of literature-connector's constraints carry over unchanged — most importantly the no-hallucination-on-failed-fetch rule: a paper appears only if its abstract was successfully retrieved.
Per-post inclusion rule: a post's section is included in the assembled post only if literature-connector returns at least one relevant paper whose abstract was successfully retrieved. If a post yields no relevant literature, omit that post's section entirely (do not write a placeholder) — but the post is still recorded in the state file (step 8), since it was genuinely evaluated.
Track, for each post evaluated this run, whether it hit (at least one relevant paper) or missed (zero). Step 8 needs the hit/miss split alongside the deferred/first-pass split from step 2.
If, after all searches, no post yielded any relevant literature, produce no post (nothing to publish), but still update the state file per step 8, and report that nothing qualified.
6. Assemble the Markdown post
Build a single Markdown document with one section per included source post. The first line must be a Markdown # H1 (it becomes the WordPress post title — publish_digest.py strips it from the body). Use today's date (2026-08-10-style) in the title.
# Daily Literature Connections — [today's date]
> Automated draft — each section links a finding from a recent lab notebook post to recent PubMed publications and preprints. **Preprints have not been peer-reviewed.**
---
## [Source post title]
**Source:** [author], [post date] · [post url]
**Finding:** [the FINDING sentence(s) you derived]
[paste the "Relevant Literature" entries from this post's literature-connector output — each with its source badge, relationship tag, title, citation line, and 2–3 sentence summary]
**Literature summary:** [paste the 2–4 sentence synthesis from the literature-connector's Summary section]
---
[repeat one section per included source post]
> Generated by the `daily-literature-post` skill · [today's date] · draft, not peer-reviewed
- Preserve
literature-connector's relationship tags (Supports / Conflicts / Adds context / Suggests next step) and source badges ([PubMed],[bioRxiv preprint — not peer-reviewed], etc.) verbatim — do not rewrite them. - Carry over the preprint caveat block from
literature-connectorat the bottom of the assembled post (the blockquote noting coverage is limited to the last 12 months and that preprints are not peer-reviewed) so the caveat travels with the published draft. Do not copyliterature-connector's full header block (query strings, retrieval counts) into the post.
7. Enforce the daily cap, decide live vs. draft, then publish
Write the assembled Markdown to a file first (so publish_digest.py can read it), e.g. digests/daily-literature-[today].md. Then run the two safety rails in this order before publishing.
7a. Daily cap — refuse a second post today
Read digests/.literature-state.json and count the entries in publish_log whose date equals today (YYYY-MM-DD). If that count is 1 or more, do not publish anything. Stop here: produce no new post, leave the state file unchanged, and report in step 9 that the skill has already published today and is refusing a second post per the one-post-per-day cap (quote the existing entry's status and URL). This is a hard stop, not a warning — it prevents duplicate posts if the skill is triggered twice in a day.
If the count is 0, continue.
7b. Authorization check — live only if explicitly authorized
Decide the publish status by inspecting AUTHORIZATION.md at the repository root:
- Read the file at
AUTHORIZATION.md(repo root). - Publish live (
--status publish) only if the file exists, is readable, and at least one line — stripped of leading/trailing whitespace — is exactly> AUTOMATED PUBLISHING: AUTHORIZED(the blockquote marker line under "Authorization marker"). A match anywhere else in the file — inline in a sentence, inside a code span, in explanatory prose — does not count, even if the characters are identical. This is deliberately stricter than a substring search: prose elsewhere in the file (including this very instruction) legitimately contains the phrase, and a substring match would treat that prose as authorization. - Otherwise fall back to draft (
--status draft): if the file is missing, cannot be read, or no line matches exactly. Do not attempt to "fix" or interpret a near-miss marker — anything other than that exact standalone line means draft.
Remember which mode you chose; you must report it in step 9.
7c. Publish
Publish with the status you determined in 7b:
# When AUTHORIZATION.md authorizes live publishing:
python3 scripts/publish_digest.py digests/daily-literature-[today].md --category auto-literature-connections --status publish
# Otherwise (default / unauthorized fallback):
python3 scripts/publish_digest.py digests/daily-literature-[today].md --category auto-literature-connections --status draft
--status draft is also the script's default, so omitting --status is equivalent to the draft command; pass it explicitly here to make the chosen mode unambiguous. The --category auto-literature-connections flag tags the post as a secondary signal for step 3. The category was created on the site on 2026-08-25, so the tag now takes effect and posts come back with it attached — the publishing token can attach an existing category but still cannot create new ones, so this depends on the category continuing to exist on the site. Step 3's title-prefix check remains the primary guard that excludes this skill's own posts; the category is now a working redundant second layer.
Confirm the command returned an HTTP 200/201 and a URL before proceeding, and note the reported post_status (publish or draft) — it should match the mode you intended. If publishing fails, report the error and do not update the state file (step 8) — the posts were not successfully connected/published, so they should remain eligible for the next run.
8. Update the state file
Update digests/.literature-state.json whenever step 5 evaluated at least one post — whether or not anything was published. A run that published nothing still learned something (which posts were evaluated, which came back empty) and must record it.
-
Set
last_run_dateto today's date (YYYY-MM-DD). -
Route every post evaluated by
literature-connectorin step 5 to exactly one ofconnected_urlsordeferred_urls, by its hit/miss result and whether it was already deferred:Hit (>=1 relevant paper) Miss (zero) First evaluation connected_urlsdeferred_urls,first_deferred= todayRetry (was in deferred_urls)connected_urls, remove fromdeferred_urlsconnected_urls, remove fromdeferred_urlsBoth
connected_urlsanddeferred_urlsare unions over their existing contents — keep all entries not touched this run, no duplicates. A URL must never appear in both lists; adding to one always means removing from the other. -
Retire stale deferrals. Any
deferred_urlsentry whosefirst_deferredis more than 14 days before today has aged out of every plausible fetch window and will never be re-fetched. Move it toconnected_urlsand drop it fromdeferred_urls, so the list cannot grow without bound. -
If a post was published, append one entry to
publish_logrecording it:{"date": "<today>", "status": "<publish|draft>", "url": "<the URL publish_digest.py returned>"}. Use the actualpost_statusfrom the publish output forstatus. This entry is what the step 7a cap reads tomorrow — and, if the skill is somehow triggered again today, what makes 7a refuse a second post. If nothing was published, leavepublish_loguntouched. -
Do not add posts excluded in steps 2–4 as logistical/already-processed/own-output (already-processed ones are already present; logistical and own-output ones were never real candidates).
-
If the file did not exist, create it now with
last_run_date= today,connected_urlsanddeferred_urlsfilled per the table above, andpublish_log= a list containing the single entry for this publish. -
Write valid JSON with the same four top-level keys (
last_run_date,connected_urls,deferred_urls,publish_log). Keep the file committed — never gitignore it.
9. Report back
In the main conversation, report one of:
- Published: the post URL, whether it went out live or as a draft, and — when it fell back to draft — the reason (no
AUTHORIZATION.md, missing/alteredAUTOMATED PUBLISHING: AUTHORIZEDmarker, or unreadable file). Also give how many source posts were included and their titles. - Refused (daily cap): the skill already published today. Report that it is refusing a second post per the one-post-per-day cap, and quote today's existing
publish_logentry (status + URL). The state file is left unchanged. - Nothing qualified today: state plainly why (no posts in the window, all posts logistical, all already processed, or no relevant literature found for any finding). Note whether the state file was updated (it is only updated when at least one post was actually evaluated by
literature-connector).
Never fabricate a published post or literature results, and never claim a post was published live when it fell back to draft — if a step produced nothing, say so.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |