post-check

Run a draft through each target platform's real constraints and hand back a pass/warn/fix verdict per check, with the fixed text supplied inline. A bundled linter does the mechanical counting; you supply the judgment.

Context

Read social-context.md at the project root (also check .agents/social-context.md) for brand voice, default platforms, and house style rules — hashtag policy, emoji policy, link conventions. If it's missing, offer to run the social-context skill first, but don't block — ask 2–3 quick inline questions and proceed:

• Which platform(s) is this draft for?
• Does the post have media, and does it have alt text?
• Any house rules on hashtags, emoji, or links?

Workflow

1. Establish targets. Confirm which platform(s) the draft is bound for. Evaluate one draft per platform separately — a draft can pass X and fail LinkedIn, and averaging verdicts helps nobody.

2. Run the linter for the mechanical checks. It computes platform-accurate character counts (X bills URLs at 23 and emoji at 2; Mastodon bills URLs at 23), the truncation point, link count and placement, hashtag count/placement/casing, engagement-bait phrasing, and formatting gotchas — deterministically, so you never eyeball a character count:

bun scripts/lint-post.mjs --platform x --file draft.txt      # one platform
bun scripts/lint-post.mjs --all --file draft.txt             # every platform
echo "$DRAFT" | bun scripts/lint-post.mjs --platform linkedin

It prints JSON — { counts, checks: [{ check, verdict, detail }], summary } — with each verdict as pass, warn, or fix, and exits non-zero when any check is a fix. node works in place of bun. Run it with --help for options, or --list for supported platforms. If neither runtime is available, count by hand against references/platform-limits.md.

1. Read the reasoning when a flag needs it. Load references/platform-limits.md for the full constraint table, the per-platform link and hashtag norms, the media/aspect rules, and the verdict-discipline definitions. Don't restate limits from memory — that reference (and the JSON the linter reads) is the source of truth.

2. Add the judgment the linter can't. The script flags mechanics; these are yours:

   • Hook. Does the first visible line carry a complete, compelling idea before the truncation point the linter reported? "So I've been thinking about someth…" is a fix — supply a rewritten opening inline.
   • Media & alt text. If media is described, is the aspect ratio right for the placement, and does the alt text actually describe the image? "image" or keyword stuffing is a fix.
   • First-comment strategy. If a link belongs in the first comment (or the draft says "link in comments"), draft that comment — one line of context plus the link.
   • Engagement bait. When the linter flags bait, replace it with an ask that requests genuine input, not a reflex.
   • Voice. Check the draft against every rule in the Voice section of social-context.md.

3. Assemble the verdict. Merge the linter's mechanical verdicts with your judgment calls into one table per platform: pass / warn / fix, a one-clause reason, and for every fix the corrected text inline. If a check doesn't apply (no media, no links), say n/a rather than dropping the row. If anything was fix, finish with a fully corrected version of the whole post for that platform.

Quality bar

Check against the real numbers in references/platform-limits.md, not a vague memory of "roughly 280". Verdict discipline:

• pass — meets the constraint; nothing more to say.
• warn — will publish but underperform or annoy; give a one-clause reason and a suggested improvement.
• fix — don't post as-is; supply the corrected text inline. A fix without corrected text is an unfinished check.

Run every check for every target platform. A silently skipped check reads as a pass.

Deliverable

Per target platform: a verdict table (check | verdict | note), each fix row followed by the corrected text inline, and — if anything was fix — a complete corrected draft for that platform, plus the drafted first comment if the strategy calls for one. The user copies the corrected draft and posts it wherever and whenever they choose — hand it back and stop.