Edit Style
The mechanism that makes the next edit better than this one. An edit gets closer to what
someone wanted when the system has seen what they actually like — not described in words, but
shown. This skill turns a dropped screenshot into a written, dated decision that
resolve-ai-edit reads at preflight.
One EDIT-STYLE.md per client project. It grows. It is the visual evidence; CREATIVE-BRIEF.md
stays the stable statement of intent, and BRAND.md stays the mechanical spec.
The rule that matters most
Read a reference for its system, never copy its composition. Extract the palette logic, the type weights and pairing, how much of the frame the graphic occupies, the materiality, the animation timing, the pacing. Then build this client's own version of that system.
Reproducing someone else's specific design is not a style decision, it is copying, and it ships under the operator's name. Say this out loud the first time a reference is supplied, then extract rather than trace.
Intake — every time a reference arrives
-
Take the file. Save it into the project's
reference/folder, named<YYYY-MM-DD>-<what-it-shows>.png. A link gets a screenshot saved the same way, plus the URL recorded — links rot, and a style decision that cannot be re-examined is a rule with no reason. If it is a video, save the frame that shows the thing, and note the timecode. -
Ask one question: what about this? A reference handed over without a reason is ambiguous — the same frame can be shown for its captions, its grade, its pacing or its layout. Do not guess. "The captions" and "the way they hold on the face" lead to entirely different extractions.
-
Look at it properly. Open it. Describe what is actually there before deciding anything: how many elements, where they sit, the type sizes relative to frame height, the colours as hex where readable, how much empty space, what the eye lands on first.
-
Extract the system, in writing. Three to six lines. Mechanical and reusable, not impressionistic. "Two-line caption, roughly 6% of frame height, heavy sans, one word in italic serif" is usable. "Clean modern look" is not.
-
Decide the scope, and say it out loud before writing anything:
Scope Test Where it goes Adopt This becomes how we do it for this client from now on EDIT-STYLE.md, as a dated decisionTry Worth one video to see EDIT-STYLE.md, markedTRY, with what would confirm itReject Shown as an example of what not to do EDIT-STYLE.mdunder Rejected, with the reasonOne-off This video only Nowhere. Do not file it -
Write it. Append to
EDIT-STYLE.md. Never rewrite an earlier entry — supersede it, and say which entry it replaces. -
Check it against what already exists. If the new reference contradicts a
BRAND.mdvalue, aCREATIVE-BRIEF.mdprohibition, or an earlier adopted entry, quote both and ask which wins. Never silently resolve it. A style record that quietly contradicts itself is worse than none, because the editor follows whichever line they read first.
What a good entry looks like
### 2026-09-18 — Caption treatment
**Reference:** `reference/2026-09-18-caption-style.png` (frame from a competitor short, 0:04)
**Shown for:** the captions
**Extracted:**
- Two lines maximum, never three. Breaks on the phrase, not the line width
- Cap height roughly 6% of frame height — large enough to read at thumb size
- Heavy sans, one word per caption in italic serif for emphasis
- Sits in the bottom third but clear of the platform UI zone
- Word-level highlight on the beat, not a fade
**Decision:** ADOPT for all short-form on this client.
**Supersedes:** 2026-08-02 caption entry (three lines, centre).
**Why:** the operator has corrected caption length on three consecutive videos.
Before an edit starts
resolve-ai-edit reads EDIT-STYLE.md at preflight. When asked what a client's style is, read
the file and answer from it — adopted entries first, then anything still marked TRY, then the
rejected list, because knowing what was ruled out and why prevents proposing it again.
If the file is long, propose a consolidation rather than doing one. Merge duplicates, retire
superseded entries, and promote the stable rules into CREATIVE-BRIEF.md where they belong as
intent. Propose. The operator approves.
What this skill does not do
- It does not edit. That is
resolve-ai-edit. - It does not invent a style. No reference, no entry. An extraction from nothing is a guess wearing a date.
- It does not carry one client's style to another. Every client's record is theirs. A treatment that worked elsewhere may be proposed with its source named, never applied silently.
- It does not touch packaging, titles or thumbnails. Different job, different skill, different repo.
Keeping this skill current
When something notable is learned, update this file at its root — not in a note or a chat log.
A correction about how a client likes their edits goes in that client's EDIT-STYLE.md. A
correction about how extraction itself should work goes here.