| name | post-ideator |
| description | Find LinkedIn post angles that can reach technical builders right now, but only when the angle is specific, differentiated, and defensible. Use when the user says "what should I post about", "pitch me ideas", "give me post angles", "find a topic", "ideate a post", or asks for fresh material without supplying a thought. Returns 3 to 5 scored idea briefs and writes shortlisted ones into ideas/YYYY-MM-DD.md. |
post-ideator
Your job is to find candidate angles for a LinkedIn post and present them as compact briefs. You are not drafting yet.
Office UI sync
This is the ideator stage — emit end once ideas/<YYYY-MM-DD>.md is written, per office-emit-end.
The selection criterion is externally relevant + sharp wedge + defensible. A hot topic alone is not enough, and neither is a personal story alone. The post needs a concrete detail, a real opinion wedge, and a reason a builder who has never heard of the owner would care.
Reach model. Reach = size of the standing audience × a sharp differentiated wedge × a front-loaded hook × a non-saturated topic. The first term is the one most often missed. "A stranger builder would care if they read it" is not the same as "a crowd is already standing around this topic." A wedge multiplies the room it lands in; it cannot conjure a room. The sharpest take on a topic only a few thousand people will ever touch still lands in a near-empty room.
The standing-audience model lives in wiki/audience.md: three tiers keyed on subject
recognizability, each with its observed impression band and its exemplar posts. Read that
page rather than reasoning about reach from intuition. Do not restate its contents here — it
gets revised as posts accumulate, and a copy in this file would go stale silently.
If wiki/ is absent (fresh clone), fall back to the top and bottom bucket sections of
bun run post-patterns and say so explicitly in your output, so the user knows the reach
scores are ungrounded that run.
Inputs to read first
- Latest briefing (required):
ls briefings/*.md | sort | tail -1 and read it. This is the canonical signal for this week.
bun run post-patterns (required): the mechanical ground truth over the archive — medians, sample sizes, anti-patterns, cooling streaks. Two cautions when reading it. The top/bottom bucket sections cover only the quartiles, so a family there can show n=2; every bucket now prints its sample size and marks thin ones too few to cite. Respect that marker. And topic_family labels come from a keyword cascade that mislabels, so treat the family axis as bookkeeping, not signal.
2.5. wiki/audience.md (required when wiki/ exists): the accumulated judgment layer. post-patterns gives you counts; the wiki gives you the conclusions those counts support, each carrying its own confidence and evidence_n. Never cite a claim whose confidence is anecdote.- Drafts dedup (required): list files in
drafts/ whose <YYYY-MM-DD> prefix is within the last 30 days, then read each one's YAML frontmatter. Collect every source_url and pitch_angle. Drop any briefing item whose URL matches, or whose angle overlaps semantically with a recent draft.
- Recent published posts (dedup):
ls posts/$(date +%Y)/*.md | sort | tail -10 and skim titles plus opening lines. Reject near-duplicates of any published post from the last 7 days.
bun run top-posts --n 10: this is the quick scoreboard for what already earned reach.
There is no cv.md in this project. Do not look for one. Do not apply lane filters tied to the user's background.
The owner's real experience
The published corpus is the only record of what the owner has genuinely done. Read it to learn their real domains (what they have built, shipped, measured, or used firsthand). Build experience-led angles only on that ground truth, and never invent a first-hand story. For each candidate, name the owner's honest entry point in an experience_hook. If a hot topic has no genuine personal angle, it can still be a strong post when it carries a sharp differentiated wedge (see the Reach model) — name the wedge in opinion_wedge and set experience_hook to "none — wedge-driven news take". Only cooling/saturated families require a firsthand artifact (see the Cooldown rule). Inventing experience is never allowed.
External-value gate (pre-filter, runs before scoring)
Before scoring any candidate, ask one question: would a builder who has never heard of the owner, and does not care about the owner's LinkedIn metrics, find this useful or interesting?
Reject outright — do not score, do not shortlist — any idea whose primary subject is the owner's own content machine: posting performance, the posting/critic/ideator pipeline, a posting losing or winning streak, follower counts, or any "here's how I run my LinkedIn" meta. This is navel-gazing. It reads as firsthand and scores well on discussion potential, but the audience is technical builders, not the owner's content ops. Three of the twelve t0 posts in wiki/audience.md are exactly this shape, including the owner's posting-streak post at 181 impressions.
source_type: build_log is valid only when the build is a real technical product or tool other builders could use, run, or learn from (BlazeDiff, Avatune, GEPA-on-Qwen, a benchmark, a migration). A build log about the post pipeline itself fails this gate.
Scoring rubric
Score each candidate angle from 0 to 2 on every axis:
heat: active attention nowspecificity: concrete number, name, or eventdifferentiation: non-obvious angle, not a recapbuilder_fit: relevant to a stranger builder (senior engineer or tooling builder) who does not know the owner — measures quality of fit, not size of audiencereach_ceiling: how large is the crowd already standing around this topic, independent of how good the wedge is. This is not "would one builder care" (that's builder_fit) — it is "how many are watching before the post exists." Do not score this from intuition. Look the subject up in wiki/audience.md, name the matching tiers[].id, name the single exemplar the candidate most resembles, and take that tier's score. Put the tier id and the exemplar in the brief's reach_ceiling line so the reasoning is auditable.
- If you think the page assigns the wrong tier for this subject, score it your way and record the disagreement in the
reach_ceiling line as disputed: <page tier> -> <your tier>, because <reason>. Never silently override a tier. Disputes get resolved against the post's actual number at retro time, which is how the tier model gets calibrated instead of just getting longer.
discussion_potential: likely to trigger real replies, not passive agreement
Only pitch items scoring >= 8/12, with these gates:
reach_ceiling is a gate. If reach_ceiling is 0, drop the candidate regardless of the total. A sharp wedge cannot rescue a small room. Every t0 post in the archive landed between 76 and 236 impressions, several of them with clean craft and a real argument. topic_family carries no reach signal — never adjust reach_ceiling on family grounds. other in particular is the classifier's residue bucket, not a family, and it contains two of the archive's top six posts.builder_fit is a gate. If builder_fit is 0 or 1, drop the candidate regardless of the total. A high total carried by heat and discussion_potential on a topic a stranger builder doesn't care about is the trap that shipped the worst posts in the archive — a company milestone announcement (76) and the owner's own posting-streak post (181).- Discussion bait doesn't count.
discussion_potential earned by personal vulnerability, confession, or meta-drama ("I failed N times", "here's what's wrong with my process") scores 0 on that axis. Count only discussion that comes from an arguable technical stake.
Carry the computed total (out of 12) through to the output — print it in the numbered brief and store it in the idea's frontmatter as score: <n>. Downstream tools (post-cycle's interactive pick) read it rather than recomputing it.
Use the briefing for heat:
- HN items: prefer high
★score and high comments.
- Lobsters items: use lower thresholds, but same logic.
- Newsletters and blogs: prefer primary sources over digest-style summaries.
- Exa fresh news: treat as a freshness tiebreaker, not the whole case.
Format: text, carousel, or decision-tree
Most ideas are single-narrative text posts. Two other formats fit specific angle shapes:
carousel — a multi-slide side-by-side comparison of 3 to 6 concrete, namable tools/options (notebook apps, runtimes, frameworks, hosting providers) where the comparison itself is the wedge, not a prose argument. Mark format: carousel only when it compares 3 to 6 specific options a reader chooses between and the payload is the per-option pros/cons, not a single linear claim.decision-tree — a single flowchart that routes the reader to a recommendation by their constraints ("which X should you use", "when to reach for Y vs Z"). Mark format: decision-tree only when the angle is genuinely a decision with 3 to 5 distinct branches keyed on real conditions (team size, latency budget, data ownership, scale), and the routing logic is the wedge.
Both still clear both gates: a comparison or decision earns reach_ceiling only when the options have a real standing audience, and builder_fit only when a stranger builder is actually making that choice.
Default to format: text. A carousel or decision-tree is a presentation choice, never a way to rescue a thin angle — score it on the same rubric and gates as any other idea.
Hard rejection rules
- Reject near-duplicates of any draft from the last 30 days.
- Reject near-duplicates of any published post from the last 7 days.
- Reject same-topic sequel posts unless there is a new artifact: experiment, code, dataset, benchmark, or strong prediction.
- Reject pure news recap angles with no opinion wedge — a flat summary that adds nothing.
- A third-party news / paper / opinion item with a sharp differentiated wedge is a valid, strong angle even without a firsthand entry point. Every
t2 exemplar in wiki/audience.md won this way. Do not reject these. What underperformed was not "news without firsthand" — it was news about a t0 subject: one vendor's product, one paper, one benchmark result. The fix is a bigger room, not a mandatory personal anchor. Set experience_hook to "none — wedge-driven news take" when there is no honest firsthand angle.
- Cooldown rule.
## Cooling families in bun run post-patterns is a trigger for review, not a gate. The report flags a family when 3 of its last 4 posts fall below its own p25, on at least 6 posts of record — but the family classifier mislabels, so confirm the streak is about a real shared topic before acting on it. When it is real, any new idea in that family needs a firsthand artifact in experience_hook: the owner's own benchmark, code, migration story, measured result, or direct usage. When the streak is a classifier artifact, say so in your output and ignore it. An unreviewed streak must not block an idea.
Recency preference
The briefing is bucketed by recency. When multiple candidates pass the rubric and the rejection rules, prefer fresher ones: Today > Last 3 days > Earlier this week.
Output format
Return 3 to 5 numbered idea briefs. Each item should be compact, but include all of these fields:
1.
angle: ...
score: 9/12
source_title: ...
source_url: ...
briefing_date: YYYY-MM-DD
why_now: ...
opinion_wedge: ...
experience_hook: <the owner's genuine first-hand entry point, or "none — wedge-driven news take">
reach_ceiling: <0 | 1 | 2> — <tier id from wiki/audience.md> · resembles <exemplar subject> · <one line naming the crowd already watching>. Append "disputed: <page tier> -> <your tier>, because ..." only when you disagree with the page.
evidence_points:
- ...
- ...
risk: generic | rehash | too niche | thin evidence
Do not draft the posts. Do not propose hashtags.
After presenting the list, write the shortlisted ideas into ideas/YYYY-MM-DD.md with YAML frontmatter entries:
---
idea_id: 2026-05-21-01
source_url: ...
source_title: ...
briefing_date: 2026-05-21
topic_family: security
source_type: news
format: text
angle: ...
score: 9
why_now: ...
opinion_wedge: ...
experience_hook: ...
reach_ceiling: 2
reach_tier: t2-universal
wiki_rev: 2026-08-17
risk: generic | rehash | too niche | thin evidence
evidence_points:
- ...
- ...
status: shortlisted
---
reach_tier is the tier id from wiki/audience.md; wiki_rev is that page's
last_revised. Together they record which version of the reach model scored this idea, so
a later retro can tell whether a miss came from the model or from the pick.
On user pick
When the user picks a brief (by number, idea id, or "do #2"), invoke the post-writer skill with the chosen idea brief as the raw material. Pass enough context for the writer to preserve the brief faithfully:
anglesource_urlsource_titlebriefing_datetopic_familysource_typeformat (text, carousel, or decision-tree; omit or text for the prose path)why_nowopinion_wedgeexperience_hookevidence_points
The post-writer handles voice, structure, and draft frontmatter. Your job ends at handoff.
If input is ambiguous
If the user said "ideate" with no further direction, pick from the strongest items in the freshest bucket that still clear the defensibility bar. Do not ask a clarifying question for a default ideation run.
