Run the full LinkedIn post pipeline end-to-end in one go — briefing (if missing), top-posts + patterns context, ideator shortlist (with an interactive pros/cons pick, scored out of 10), writer draft, critic verdict. Use when the user says "full cycle", "run the pipeline", "do the whole flow", "post-cycle", "give me a post end to end", or wants a single command that replaces invoking topics-briefing, post-ideator, post-writer, and post-critic separately. Optionally accepts an idea number to skip the pick and go straight from a freshly generated shortlist.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Run the full LinkedIn post pipeline end-to-end in one go — briefing (if missing), top-posts + patterns context, ideator shortlist (with an interactive pros/cons pick, scored out of 10), writer draft, critic verdict. Use when the user says "full cycle", "run the pipeline", "do the whole flow", "post-cycle", "give me a post end to end", or wants a single command that replaces invoking topics-briefing, post-ideator, post-writer, and post-critic separately. Optionally accepts an idea number to skip the pick and go straight from a freshly generated shortlist.
post-cycle
Orchestrates the post pipeline end-to-end so the user does not run each skill by hand. You are the conductor. The underlying skills (topics-briefing, posts-postmortem, post-ideator, post-writer, post-critic, post-image, post-carousel, post-flowchart) own their logic. Do not duplicate their rules. Do not skip them.
Inputs
args (optional) carries two independent tokens, in any order:
An integer 1..5 picking a specific idea from the freshly generated shortlist, no prompt. If absent, step 4 asks the user to pick interactively — see below.
A format token — text, carousel, or decision-tree — the requested format for this run. If absent, the format is whatever the chosen idea carries (default text).
Examples: /post-cycle carousel (interactive pick among carousel-fit ideas), /post-cycle 2 (idea 2 directly, no prompt, its own format), /post-cycle 2 decision-tree (idea 2, forced to a decision-tree, no prompt).
When a format token is present it constrains the run: the ideator is asked to surface and prioritize angles that fit that format (step 3), the pick prefers an idea already tagged with it (step 4), and the visual step routes accordingly (step 7). A requested format that cannot honestly fit any shortlisted angle is a stop condition, not something to force onto a mismatched idea.
Workflow
Run the steps in order. Each step produces a concrete artifact on disk. Do not stop between steps unless a step truly fails.
1. Briefing (only if missing)
Check if today's briefing exists:
test -f "briefings/$(date +%Y-%m-%d).md" && echo present || echo missing
present: skip to step 2 — do not regenerate.
missing: invoke the topics-briefing skill. Wait for it to finish writing briefings/$(date +%Y-%m-%d).md.
Briefings are dated. Yesterday's file does not count as today's; only same-day matches.
1.5. Postmortems (only if stale)
The writer and critic need failure context, not just success context. Postmortems live in retros/postmortems/ and are slow-changing, so they do not regenerate every cycle.
Check freshness:
test -d retros/postmortems && \
find retros/postmortems -type f -name '*.md' -mtime -14 -print | head -1
If the command prints at least one path, postmortems are fresh enough — skip.
If the directory is empty or the newest postmortem is older than 14 days, invoke the posts-postmortem skill. It writes one file per underperforming post to retros/postmortems/.
If posts-postmortem fails (corpus too small, etc.), surface the warning and continue. A missing postmortem corpus is not a pipeline failure.
1.6. Retro sweep (pending, last 3 months only)
The 72-hour retro has no trigger of its own — a post is drafted here, but it goes live and matures after this cycle ends, so nothing comes back for it. Sweep for that debt at the top of every cycle: any published post old enough to have stable numbers but still missing its retro.
A post is a retro candidate when all hold:
It has a local draft in drafts/ (post-retro reads the draft; posts without one can't be retro'd and are skipped).
It is published — a matching file exists under posts/, which is also where the 72h metrics come from.
Its posted_at is more than 72 hours ago (numbers have settled) and within the last 3 months (older posts are out of scope — do not retro them).
No retros/<YYYY-MM-DD-slug>.md exists yet.
List the candidates:
for d in drafts/*.md; do
base=$(basename"$d" .md) # YYYY-MM-DD-slug
post=$(find "posts/${base:0:4}" -maxdepth 1 -name "${base:5}*.md" 2>/dev/null | head -1)
[ -z "$post" ] && continue# not published yettest -f "retros/$base.md" && continue# already retro'd
posted=$(grep -m1 '^[[:space:]]*posted_at:'"$post" | sed "s/^[^:]*: *//;s/[\"']//g")
echo"$base | posted=$posted"done
Archive post slugs are often longer than the draft slug (the scraper keeps more of the first line), so match by prefix with find rather than an exact path — an exact match would miss every published post. The retro filename keys off the draft slug ($base), the same name post-retro writes, so the existence check stays consistent.
For each line printed, compare posted to now: skip if newer than 72h (not ready) or older than 3 months (out of scope); otherwise invoke the post-retro skill with that YYYY-MM-DD-slug as its argument. Run them one at a time. If a retro fails (missing metrics, etc.), surface the warning and continue — a stuck retro never blocks the rest of the cycle.
1.7. Absorb retro lessons into the wiki
A retro's conclusion reaches no downstream consumer on its own — post-patterns surfaces only decision and summary, capped at three summaries per decision, so lessons expire silently as retros accumulate. Anything the sweep just wrote is still unabsorbed.
If the count is above zero, invoke wiki-curator in ingest mode. It absorbs one lesson at a time, reconciles it against the existing pages, appends to wiki/log.md, and flips the flag. Soft failure: if it errors, surface the warning and continue — an unabsorbed lesson does not block a draft.
Skip when the count is zero.
2. Archive context
Run both scripts and keep their output in context for the ideator and critic:
bun run top-posts --n 10
bun run post-patterns
bun run wiki lint
If top-posts or post-patterns fails, surface the error and stop — the ideator and critic both depend on post-patterns.
wiki lint is advisory here: report any error it prints and carry on. A stale wiki page still beats no wiki page, and the ideator falls back to the pattern report when wiki/ is missing entirely.
Two cautions when passing this context on. Bucket sections cover only the quartiles and print their own sample size — respect the too few to cite marker. And cite a fault only from ## Validated anti-patterns; the ## Tested and discredited list has been checked against the corpus and does not mark weaker posts.
3. Ideate
Invoke the post-ideator skill. It will:
Read the latest briefing.
Use the post-patterns and top-posts context already gathered.
Dedup against recent drafts and published posts.
Emit 3 to 5 scored idea briefs and write shortlisted entries into ideas/$(date +%Y-%m-%d).md.
If args carried a format token (carousel or decision-tree), tell the ideator to surface and prioritize angles that fit that format and tag them with it. The format never lowers the bar — a forced format still has to clear the rubric and gates. If the ideator can produce no idea that honestly fits the requested format, stop and tell the user (do not bend a mismatched angle into the wrong shape).
Do not draft yet. Wait for the ideator to finish.
4. Pick an idea
If args has a number N, pick the Nth idea from the just-written ideas/$(date +%Y-%m-%d).md (order = order in file) — an explicit number is a direct override, skip the interactive pick below entirely.
If the pick is out of range, stop and ask the user which idea to use.
Otherwise (no number in args), run the interactive pick below.
Interactive pick. Read every shortlisted idea's score (out of 12), opinion_wedge, reach_ceiling, and risk from ideas/$(date +%Y-%m-%d).md. Rank by score descending; if a format token was requested, rank ideas already tagged with that format first. Ask the user to choose with AskUserQuestion, one option per idea:
Cap at 4 options (the tool's hard max). With 5 shortlisted ideas, offer the top 4 by the ranking above and say in the question text that a 5th exists and can be reached by typing its number into "Other".
Option label: rank + display score, e.g. #1 · 8/10. Convert the ideator's score (out of 12) to the friendlier out-of-10 shown here: round(score * 10 / 12).
Option description: compress the brief's own fields into pros and cons — do not invent new judgment beyond what the brief already says. Pros: the opinion_wedge, plus the reach_ceiling note when it's 2 (mainstream/hot topic — a bigger room). Cons: the risk field, plus the reach_ceiling note when it's 1 (capped at a sub-community). If the requested format differs from the idea's own format, add one clause noting the conversion (e.g. "would convert from text to carousel"). Keep the whole description under ~50 words.
Question header: Pick idea. Question text: Which idea should we draft next?
Map the answer back to its idea_id. If AskUserQuestion cannot be presented (no interactive channel available), fall back to the highest-ranked idea and note the fallback in the step-1 status line.
If a format was requested and the chosen idea's format differs, set the idea's format to the requested value via Edit — but only if the angle genuinely supports that shape (a real 3-to-6-option comparison for carousel; a real 3-to-5-branch decision for decision-tree). If it cannot fit, stop and tell the user the picked angle does not suit the requested format.
Mark the chosen idea's status in ideas/$(date +%Y-%m-%d).md to approved via Edit. Leave the others as shortlisted.
5. Draft
Invoke the post-writer skill with the approved idea brief. Pass through every field the writer expects: angle, source_url, source_title, briefing_date, topic_family, source_type, format, why_now, opinion_wedge, evidence_points. The writer drafts in the brief's format (prose for text, the per-slide outline for carousel, the routing tree for decision-tree).
The writer saves to drafts/$(date +%Y-%m-%d)-<slug>.md. Capture that path for the next step.
6. Critique
Invoke the post-critic skill with the draft path from step 5. The critic will read the idea brief, the patterns report, and the draft, then emit either an approval scorecard or a rewrite plan.
If the critic rejects, stop here. Do not generate a concept image for a draft that is not going to publish.
7. Visual concept (only if critic approved)
Read the approved draft's format and branch to the matching visual skill:
format: text (or absent) → post-image. Always square; there is no size to pick.
format: carousel → post-carousel. Always square 1080×1080; there is no size to pick. Writes an index concepts/<date>-<slug>/prompt.md plus slide-NN.md files.
format: decision-tree → post-flowchart. Always square; there is no size to pick. Writes a single concepts/<date>-<slug>/prompt.md.
Whichever runs will:
Reason over the draft to produce the right visual (one metaphor for a text post; an intro/closing metaphor plus per-tool icons and a shared visual system for a carousel; a clean routing flowchart for a decision-tree).
Write the prompt(s) under concepts/<date>-<slug>/, with prompt.md as the file concept_path points at.
Update the draft frontmatter with concept_path: concepts/<…>/prompt.md.
Call bun run generate-image concepts/<date>-<slug> to actually render the prompt(s) with OpenAI's gpt-image-2 into the gitignored images/<date>-<slug>/ folder, when OPENAI_API_KEY is set.
Print the prompt(s) either way so the user can paste them into an image tool.
Surface the style choice rather than taking the default silently — style is the highest-leverage visual variable. If the visual skill fails to build or save the prompt itself, surface the error and stop — the draft is already saved and approved, and the user can re-run /post-image <draft> (text), /post-carousel <draft> (carousel), or /post-flowchart <draft> (decision-tree) manually. A missing OPENAI_API_KEY or a failed generation call is not this kind of failure — the prompt is still saved, so treat it as a soft skipped/failed note and continue to Output.
Output
Print, in order:
A one-line status per step: briefing: created | reused, postmortems: refreshed | fresh | skipped, retros: N swept | none pending, ideation: N briefs, pick: interactive | direct (arg N) | fallback, chosen: <idea_id>, format: text | carousel | decision-tree, draft: <path>, critic: approved | rejected, concept: <path> | skipped, image: generated | skipped (no OPENAI_API_KEY) | failed | N generated, K failed.
The full draft text (so the user can read it without opening the file).
The critic's verdict block verbatim.
If the concept was generated, the visual skill's output block verbatim, including its Image:/Images: line — the single post-image prompt for a text post, the delimited post-carousel slide prompts (--- SLIDE NN ---) for a carousel, or the single post-flowchart prompt for a decision-tree. When generation succeeded, tell the user the PNG(s) are already sitting in images/<date>-<slug>/; when skipped or failed, the prompt is still there to paste manually.
If the critic rejected the draft, stop after step 3. Do not auto-revise and do not generate a concept image. The rewrite plan is the user's call.
Failure handling
Briefing step fails: report the failure and stop. The rest of the pipeline depends on a current briefing.
bun run top-posts or bun run post-patterns fails: report and stop.
Ideator returns zero passing briefs: stop and tell the user the shortlist was empty. Do not invent an angle.
Writer fails to save the draft: stop. Do not retry without the user.
Critic rejects: this is not a pipeline failure. Print the rewrite plan and end normally.
What this skill does not do
It does not run post-retro. That happens 72 hours after publishing.
It does not publish to LinkedIn.
It does not bypass the critic. A rejected draft stays a draft.
It does not edit posts/. Those are scraped output.
When to use
Trigger phrases:
"run the full cycle"
"do the whole flow"
"post-cycle"
"give me a post end to end"
"pipeline a post"
"ideate, draft, and critique"
If the user already has a draft and just wants it critiqued, use post-critic directly. If they already picked an idea, use post-writer directly. This skill is for the cold start.
