Creative preflight: an offline editorial checklist for draft posts

Status: experimental, human review only, not yet in a released package. This command does not know any network’s ranking weights. It does not predict views, check whether sources are true, measure retention or claim any uplift.

Run it

reachweaver creative-preflight --file examples/creative-preflight.example.json
reachweaver creative-preflight --file examples/creative-preflight.example.json --json

python3 -m reachweaver creative-preflight ... does the same from a source checkout. With the standalone Claude Code skill, run python3 <skill-dir>/scripts/run.py creative-preflight --file <skill-dir>/assets/creative-preflight.example.json, like every other command.

The input is one local JSON file that you control. Claude Code can help you prepare the drafts, but check every source, permission, claim and privacy question yourself before you share content with an AI provider or publish it.

Exit codes: 0 when the brief was read and checked; 2 with one ERROR: line when the file is missing, larger than 256 KB, a symlink, not UTF-8, not valid JSON or does not match the schema. The error names the field at fault.

The brief

schema_version is reachweaver.creative-preflight.v1. Required fields:

• topic: up to 160 characters;
• audience: a description and 1-10 keywords;
• goal: awareness, discussion or qualified_visits;
• candidates: 1-6 drafts, each with a unique id, a platform from the profile-audit registry, a format and the text.

Optional fields:

• facts: up to 10 entries, each with an id, a claim and an https:// source_url; a draft names the facts it relies on in evidence_refs;
• per draft: title, cta, alt_text and a true/false captions.

No URL is visited. Unknown fields, unknown platforms, undefined fact IDs and credential-like strings are refused.

What the report means

Seven disclosed weights (20/20/20/15/10/10/5) measure how complete the editorial checklist is. They do not measure algorithm compatibility or the chance of reach. The checks are:

1. a declared audience term appears in the draft;
2. the opening line is 5-26 words and makes no hype promise;
3. the draft names owner-supplied facts;
4. the format is one the registry lists for that platform;
5. the topic appears in the draft;
6. captions (video) or alt text (images) are declared;
7. the declared call to action appears in the draft.

A check that cannot be judged is unknown, not a failure, and assessed_weight shows how much was judged. For example, an opening in a script written without spaces (Chinese, Japanese, Thai) is unknown, because word counts mean nothing there.

A full score never proves that the claims are true, that the creative is original or that a platform will promote the post. The lexical checks can miss multilingual or semantic relevance. Compare individual findings, not scores between networks.

The command reuses profile_audit_registry.json and the seo-content-check rules for repeated or heavy hashtags, text-length limits and near-identical wording across drafts. The three brainstorming questions are prompts for your own writing, not ready-to-post text.

Voice findings (optional)

--voice-profile FILE adds advisory voice_findings per draft from a profile made by voice-profile: features outside the usual range of your own posts and stock phrases from a heuristic style list (not an AI detector). The text report lists the warn ones; the checklist points and the exit code do not change. See voice check.

Human-only destinations

Reddit and Hacker News are human-only, and DEV is human-authored only. A draft for one of them is still checked, with a human_only note. When a brief contains one, the report has no brainstorming prompts at all: write that text yourself under the community’s rules.

Where it fits

Run it before campaign-preview, claims-check and repeat-check. Publishing is still controlled by exact-item human approval, the live gates, caps and the other safety rules. Platforms without an adapter stay manual-only. The command opens no database and makes no network request. It calls no model, connects no account and writes no file.