Run your blog like a queue, not a guessing game. This skill is the orchestration layer on top of blog-topic-research and seo-blog-writer: it keeps an evidence-backed backlog, picks the next topic so your corpus drifts toward the cluster + format mix you defined, schedules posts into a rolling daily cadence, and marks them done — autonomously, two posts a day if you want. It scores every pick by how far each topic cluster is from its target weight (so you never end up with 40 troubleshooting posts and no comparisons), reconciles the backlog against what's actually live on your CMS, and auto-refills by calling blog-topic-research when the queue runs dry. Platform-agnostic: it drives whatever publish adapter seo-blog-writer is configured for (Ghost, WordPress, static site). Built for founders, indie hackers, and content teams who want a hands-off long-tail SEO pipeline that publishes on a schedule and stays balanced. Trigger when the user says: 'what should I write next', 'pick the next topic', 'schedule the nex
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Run your blog like a queue, not a guessing game. This skill is the orchestration layer on top of blog-topic-research and seo-blog-writer: it keeps an evidence-backed backlog, picks the next topic so your corpus drifts toward the cluster + format mix you defined, schedules posts into a rolling daily cadence, and marks them done — autonomously, two posts a day if you want. It scores every pick by how far each topic cluster is from its target weight (so you never end up with 40 troubleshooting posts and no comparisons), reconciles the backlog against what's actually live on your CMS, and auto-refills by calling blog-topic-research when the queue runs dry. Platform-agnostic: it drives whatever publish adapter seo-blog-writer is configured for (Ghost, WordPress, static site). Built for founders, indie hackers, and content teams who want a hands-off long-tail SEO pipeline that publishes on a schedule and stays balanced. Trigger when the user says: 'what should I write next', 'pick the next topic', 'schedule the next N posts', 'editorial calendar status', 'add this to the backlog', 'mark this post done', or any variant of running a recurring publish cadence.
The scheduler and topic-picker that sits on top of the other two skills. blog-topic-research fills a backlog with evidence-backed topics; seo-blog-writer turns one topic into a published post. This skill decides which topic to write next and when to publish it, so the corpus drifts toward the content mix you defined instead of whatever you felt like writing that day.
/blog-editorial-calendar next [N] # pick + write + auto-schedule into the next slots
/blog-editorial-calendar next [N] --draft # pick + write, leave as a draft (no schedule)
/blog-editorial-calendar next [N] <slot>[, <slot> ...] # pick + write + schedule each explicit slot
/blog-editorial-calendar status
/blog-editorial-calendar add "<topic>" --cluster <c> --format <f> [--priority N] [--notes "..."]
/blog-editorial-calendar mark-done <id-or-slug>
/blog-editorial-calendar sync # reconcile the backlog against what's live on your CMS
next always writes. With no flags it schedules each post into the next free slot of the rolling cadence; with --draft it writes but leaves the post as a draft; with explicit slot args it uses the named windows. The author and target platform come from config.json (see below) — this skill never overrides them.
Configuration — config.json
Everything site-specific lives in skills/blog-editorial-calendar/config.json, not in this file. Edit it once for your blog; the reference code below reads it. Example:
cluster_targets — your topic buckets and the share of the corpus each should hold. The names are yours; the example above is a placeholder. They must sum to ~1.0. This is the primary balancing axis.
format_targets — post-shape mix (universal; the defaults above are sensible for a how-to blog). Secondary axis.
cluster_tag_map — maps each cluster to the CMS tag(s) sync uses to classify live posts.
cadence — how many posts/day and the UTC windows they fall into.
target_platform / author — passed straight through to seo-blog-writer.
Status model — the dedup gate
The backlog is the single source of truth for what topics are taken. Every topic flows through three statuses:
Status
Meaning
Set by
queued
Researched and ready to write
add, or blog-topic-research --append
in-progress
Writer has started; publish not yet complete
seo-blog-writer at start (auto-creates the row if missing)
done
Uploaded to the CMS in any state — draft, scheduled, or published
seo-blog-writer the moment the publish call returns 2xx
Plus killed for topics intentionally dropped (kept so research doesn't re-add them).
This vocabulary is the entire dedup mechanism. The picker filters on status == "queued" only — it never needs a title-similarity match at pick time. As long as the writer locks at start and flips at publish, the picker cannot pick a topic that's already in flight or already shipped. If the writer crashes mid-flow, revert the orphan row to queued (with a last_error note) so the topic re-enters the pool.
Data files
backlog.json lives at skills/blog-editorial-calendar/backlog.json. Schema for a queued entry:
{"topics":[{"id":"self-host-on-a-vps","topic":"Self-hosting <product> on a $5 VPS: complete 2026 setup walkthrough","cluster":"core-product","format":"how-to-automate","priority":1,"status":"queued","tags":["Tutorial","Product"],"notes":"anchor post - long, link-worthy","added_at":"2026-01-04","published_slug":null,"published_at":null,"research_proof":{"demand_signals":[{"type":"github_issue","url":"https://...","evidence":"verbatim issue title"},{"type":"keyword_data","url":"dataforseo:labs/keyword_overview","evidence":"\"self host <product> vps\" search_volume=210/mo, keyword_difficulty=12"}],"primary_sources":["https://docs..."],"keywords":["primary keyword","lsi 1","lsi 2"],"problem_summary":"1-2 factual sentences in writer-voice. Empty string if not derivable.","confirmed_fixes":[{"kernel":"short fix phrase","source":"https://..."}],"version_context":"v1.65+ or null","question_variants":["paraphrase 1","paraphrase 2"]}}]}
The research_proof blob is written by blog-topic-research --append and carries the writer a vetted scaffold so its research step telescopes. It is stripped on done-flip (see mark-done) so the file stays small across hundreds of posts.
Before running the requested command, refresh the local index of what's live on your CMS, then reconcile the backlog against it. The writer's platform adapter already knows how to list posts — reuse that. Write the result to state/published-index.json (gitignored): one row per post with slug, title, status, published_at, tags. All three CMS states — published, scheduled, draft — count as "this topic is taken".
# Platform-neutral: ask seo-blog-writer's configured adapter to list posts.# (Ghost example — substitute your platform's list endpoint.)# GET /ghost/api/admin/posts/?limit=all&fields=id,slug,title,status,published_at&include=tags
After the index refreshes, run the sync reconciliation. If the list call fails (credentials missing, network error), print sync skipped: <reason> and continue against the stale index — status is fine on stale data, and the writer's start-of-run lock catches any race the index would have.
Write-through: when the writer flips a backlog row to done at publish time, it should also upsert the new post into state/published-index.json so the index stays consistent without waiting for the next preflight.
Else if --draft is present → draft mode (write, leave as a draft).
Otherwise → auto-schedule mode (the default).
Input
Result
next
1 post, scheduled in the next free slot
next 4
4 posts across the next slots
next 1 --draft
1 post, left as a draft
next today evening
1 post in tonight's evening window
next 3 mornings
3 posts in the next 3 morning slots
Never silently fall back from slot mode to auto-schedule. If a slot keyword is present and date parsing fails, abort with a clear error.
Auto-schedule mode (default)
Walks the cadence forward from the latest scheduled (or, if none, published) post. With posts_per_day: 2 and the two default windows, slots alternate morning → evening → morning → evening, every day. For each of N picks:
The writer flips the row to done, sets published_slug + published_at when the publish call returns 2xx. No mark-done needed in the happy path.
On failure the writer reverts the row to queued with last_error and the loop stops — don't grind through the backlog while one topic is broken.
Cadence reference parser
Prints N ISO-8601 UTC timestamps, chronological, one per line.
python3 - "<N>" <<'PY'
import datetime, json, pathlib, random, sys
N = int(sys.argv[1])
base = pathlib.Path("skills/blog-editorial-calendar")
cfg = json.loads((base / "config.json").read_text())
SLOTS = [tuple(s) for s in cfg["cadence"]["slots_utc"]] # [(sh,sm,eh,em), ...]
now = datetime.datetime.now(datetime.timezone.utc)
data = json.loads((base / "backlog.json").read_text())
done_pub = [t for t in data["topics"]
if t.get("status") == "done" and t.get("published_at")]
def window(day, idx):
sh, sm, eh, em = SLOTS[idx]
start = datetime.datetime(day.year, day.month, day.day, sh, sm, tzinfo=datetime.timezone.utc)
end = datetime.datetime(day.year, day.month, day.day, eh, em, tzinfo=datetime.timezone.utc)
return start, end
if done_pub:
latest = max(t["published_at"] for t in done_pub)
last_dt = datetime.datetime.fromisoformat(latest.replace("Z", "+00:00"))
# find which slot the last post fell in, then advance one
idx = max((i for i, (sh, *_ ) in enumerate(SLOTS) if last_dt.hour >= sh), default=0)
if idx + 1 < len(SLOTS):
cur_day, slot_idx = last_dt.date(), idx + 1
else:
cur_day, slot_idx = last_dt.date() + datetime.timedelta(days=1), 0
else:
cur_day, slot_idx = now.date(), 0
stamps, guard = [], 0
while len(stamps) < N and guard < 365 * len(SLOTS) + len(SLOTS):
guard += 1
start, end = window(cur_day, slot_idx)
if end > now:
floor = max(start, now + datetime.timedelta(minutes=2))
span = max(0, int((end - floor).total_seconds()))
stamps.append(floor + datetime.timedelta(seconds=random.randint(0, span)))
slot_idx += 1
if slot_idx >= len(SLOTS):
slot_idx = 0
cur_day += datetime.timedelta(days=1)
if len(stamps) < N:
sys.exit(f"could not fit {N} slots within a year")
for t in stamps:
print(t.isoformat().replace("+00:00", "Z"))
PY
Draft mode
Same loop, but invoke /seo-blog-writer "<topic>" --draft instead of --publish-at. No slot timestamp is reserved.
Slot mode
Parse explicit windows (today evening, 7 May morning, mornings, etc.), sort timestamps ascending, then run the same pick → write loop per slot. If a singular slot ends in the past at run time, abort before writing anything; plural slots auto-skip past windows.
Scoring
Two-stage: cluster first (primary balancing axis), then format inside the chosen cluster (secondary axis). Both targets come from config.json.
python3 - "<N>" <<'PY'
import json, pathlib, sys
N = int(sys.argv[1])
base = pathlib.Path("skills/blog-editorial-calendar")
cfg = json.loads((base / "config.json").read_text())
CLUSTER_TARGET = cfg["cluster_targets"]
FORMAT_TARGET = cfg["format_targets"]
data = json.loads((base / "backlog.json").read_text())
topics = data["topics"]
done = [t for t in topics if t["status"] == "done"]
queued = [t for t in topics if t["status"] == "queued"]
c_counts = {c: 0 for c in CLUSTER_TARGET}
f_counts = {f: 0 for f in FORMAT_TARGET}
for t indone:
c_counts[t["cluster"]] = c_counts.get(t["cluster"], 0) + 1
f_counts[t["format"]] = f_counts.get(t["format"], 0) + 1
total = sum(c_counts.values())
c_deficit = {c: CLUSTER_TARGET[c] - (c_counts[c] / max(total, 1)) for c in CLUSTER_TARGET}
f_deficit = {f: FORMAT_TARGET[f] - (f_counts.get(f, 0) / max(total, 1)) for f in FORMAT_TARGET}
recent = sorted([t for t indoneif t.get("published_at")],
key=lambda t: t["published_at"], reverse=True)
last_cluster = recent[0]["cluster"] if recent else None
last_format = recent[0]["format"] if recent else None
picks = []
for _ in range(N):
c_score = {c: c_deficit[c] - (0.10 if c == last_cluster else 0) for c in CLUSTER_TARGET}
avail = {c: [t for t in queued if t["cluster"] == c and t not in picks] for c in CLUSTER_TARGET}
chosen = None
for c in sorted(CLUSTER_TARGET, key=lambda c: c_score[c], reverse=True):
if not avail[c]:
continue
def key(t):
fmt = t["format"]
return (-(f_deficit.get(fmt, 0) - (0.05 iffmt == last_format else 0)), t.get("priority", 100))
chosen = sorted(avail[c], key=key)[0]
breakif not chosen:
break
picks.append(chosen)
last_cluster, last_format = chosen["cluster"], chosen["format"]
print(f"Current mix ({total} done):")
for c in CLUSTER_TARGET:
pct = (c_counts[c] / max(total, 1)) * 100
print(f" {c:16s} {c_counts[c]:3d} ({pct:4.1f}% / target {CLUSTER_TARGET[c]*100:.0f}%) deficit {c_deficit[c]*100:+5.1f}pp")
print(f"\nNext {len(picks)} pick(s):")
for i, pk in enumerate(picks, 1):
print(f" {i}. [{pk['cluster']:14s}/{pk['format']:14s}] {pk['topic']} (id={pk['id']})")
PY
The -0.10 (cluster) and -0.05 (format) penalties stop the picker from stacking the same cluster/format two posts in a row. priority is only a tiebreaker.
Backlog refill
If scoring returns no pick — queued is empty across every cluster — pause and invoke /blog-topic-research 10 --append, which validates candidates and prompts for confirmation before writing them as queued rows. Then re-run scoring and continue. Refill fires at most once per next N invocation; if the queue empties again in the same call, halt with a shortfall message rather than re-prompting. This is a fallback, not a routine pre-fetch — keep the queue ahead with proactive research runs.
Command: status
Print queued / in-progress / done counts per cluster and format, current-vs-target mix, age of the oldest queued item, and count of stale (>30-day) queued items. Reuses the counting code from scoring.
Command: add
Validate cluster is a key in config.json, format is in the format set, and id is unique (kebab-case from the topic, stop-words stripped, ≤60 chars). Append and save. Refuse if any existing row already has that id.
import re
STOP = {"the","a","an","for","with","in","to","of","on","and","or","is","are"}
slug = "-".join(t for t in re.findall(r"[a-z0-9]+", topic.lower()) if t notin STOP)[:60].rstrip("-")
Command: mark-done
In the happy path the writer flips the row automatically. This is the manual fallback (post made directly in the CMS, or a writer that crashed after publish but before the flip). Find by id or published_slug, set status = "done", and strip the heavy fields so the backlog stays lean.
Done-row schema (only these kept): id, topic, cluster, format, tags, status, published_slug, published_at. Removed on flip: priority, notes, added_at, research_proof, last_error.
python3 - "<id-or-slug>""<actual-slug>" <<'PY'
import json, pathlib, datetime, sys
key, slug = sys.argv[1], sys.argv[2]
p = pathlib.Path("skills/blog-editorial-calendar/backlog.json")
data = json.loads(p.read_text())
KEEP = {"id","topic","cluster","format","tags","status","published_slug","published_at"}
hit = next((t for t in data["topics"] if t["id"] == key or t.get("published_slug") == key), None)
if not hit:
sys.exit(f"not found: {key}")
hit["status"] = "done"
hit["published_slug"] = slug
hit["published_at"] = hit.get("published_at") or datetime.datetime.now(datetime.timezone.utc).isoformat().replace("+00:00","Z")
for k in list(hit):
if k not in KEEP:
del hit[k]
p.write_text(json.dumps(data, indent=2, ensure_ascii=False) + "\n")
print(f"marked done: {hit['id']} -> {slug}")
PY
Command: sync
Reconcile the backlog against state/published-index.json (refreshed in preflight). Auto-runs before every command.
For each live post, match a backlog row by published_slug, else by Jaccard token overlap ≥0.7 against the title (slug match is authoritative).
Matched + stale (not done, or slug differs) → set done, fill published_slug + published_at, strip heavy fields.
Matched + killed → surface, don't auto-revive.
Unmatched live post → auto-import as a done row, deriving cluster from the primary tag (via cluster_tag_map), format heuristically from the title (vs → x-vs-y; what is → what-is; fix/error → how-to-fix; connect/integrate → how-to-connect; default → how-to-automate). If no tag matches a cluster, set cluster: "?" and surface for manual fix.
Backlog done rows whose slug is in none of the live sets → surface as missing (probably deleted in the CMS).
Does not draft posts. It delegates the whole writing pipeline to seo-blog-writer. It's a scheduler around that, not a replacement.
Does not pick the platform or author. Both come from config.json and pass through to the writer untouched.
Does not weight format above cluster. Cluster mix is primary; format is secondary, applied within the chosen cluster.
Does not run more than one loop at a time. The in-progress status is the lock; a second next while one is in flight should detect it and refuse.
Does not touch git. Treat each add / mark-done / next like a normal source edit; you commit.
Does not refill silently. When the queue empties, refill calls blog-topic-research, which shows candidates and asks before writing them. Fires at most once per next N.
One-command summary
/blog-editorial-calendar next [N] # pick -> write -> schedule into the next cadence slots
/blog-editorial-calendar next [N] --draft # pick -> write -> leave as a draft
/blog-editorial-calendar next [N] <slot>[, <slot>...] # pick -> write -> schedule each explicit slot
Load backlog.json + config.json. Bucket queued/done by cluster and format.
Compute deficit vs your target weights.
Penalize the most-recently-done cluster (−0.10) and format (−0.05) to avoid stacking.
Pick the highest-scoring cluster with queued items, then the topic whose format has the largest deficit; priority breaks ties. If the queue is empty, refill once via blog-topic-research.
Compute the next slot timestamp (auto-schedule), use the explicit slot (slot mode), or skip timestamping (draft mode).
Lock to in-progress, invoke seo-blog-writer; it flips the row to done on a 2xx publish. On failure it reverts to queued and the loop halts.
For N>1, repeat with the just-picked cluster/format treated as "most recent" and advance the cadence.
