# Changelog — Drive Social creative pipeline Three skills, one workflow: Codex `$brand-standards` → Claude Cowork `/creative-strategy` → Codex `$meta-ad-creative-generator`. Versions are per skill; the handoff contract between them is versioned separately (`drive_social_handoff`). | Skill | Current version | Installed | |---|---|---| | `$brand-standards` (Codex) | 4.0 | 2026-09-11 | | `/creative-strategy` (Claude Cowork) | 4.3 | 2026-09-11 | | `$meta-ad-creative-generator` (Codex) | 4.1 | 2026-09-11 | | Handoff contract | v1 out of Step 1 · v2 out of Step 2 | 2026-09-11 | --- ## v4 — 2026-09-11 · "Ideas, not templates" **Why.** The v3 skills produced visually different ads without producing better ads. Step 2 was assigning layouts (format families, color modes, rotation quotas, a mandatory no-photo card per bucket) and Step 3 was executing them, so weak design decisions were being specified upstream and rendered faithfully downstream. A Codex self-audit and a Claude review agreed on the same diagnosis. v4 moves visual decisions to the step that can see the pixels, gives every concept a visual idea instead of a layout, and adds a real visual-review gate before anything is presented. The workflow also changes shape: the client approves or rejects rendered ads, approved ads go live, and rejected ones are remade as new options in a follow-up — per-ad editing is retired. ### `/creative-strategy` — v3 → v4.3 (Claude Cowork) Added - **Concept count is asked every run** — "5 or 10? (or type a number)" plus a reserve count. The number is a ceiling, not a quota; weak concepts are cut, never padded. - **Visual idea per concept.** Three new required fields: *Visual idea*, *Message–image relationship*, *Why this client*. A layout description no longer counts as a concept. - **Per-concept asset plan** — Keep real / Enhance / Generate / Must accomplish — replacing the bucket-level asset mode as the operative instruction (the mode survives as a summary only). - **Section 4: Resolved production brief.** One place listing the rules that govern the run, each tagged `client-approved requirement`, `strategist-approved direction`, `verified fact`, `proposed recommendation`, or `unresolved question`. Overrides quote the exact Appendix A rule they supersede. - **Reserve ideas per bucket**, marked NOT AUTHORIZED FOR PRODUCTION, held for replacements. - **Step 5: Replacement round** (`/creative-strategy replace`). Rejections are classified as *replace* (new idea, from the reserve first), *revise* (same idea, new execution), or *correct* (targeted fix) — proposed by the skill, confirmed by the strategist. Approved concepts get a `preserve` status and are otherwise byte-for-byte unchanged, verified by diffing against an archived copy of the prior round. - **Visual benchmarks** — `brand/benchmarks/approved/` and `rejected/` with one-line notes; if none exist, the first client review establishes them. - **Rejection scope.** Every client rejection is recorded as `execution`, `concept`, `campaign`, or `brand`; only the last two become restrictions. - **Explicit restrictions** carried as a separate mandatory list with source, date, and scope, distinct from suggested treatments and from anything unspecified. - Bucket rules that were previously repeated by hand: hero-product businesses, restaurants by occasion, shared-menu multi-location clients, both tonal registers by default, outside research as creative substance, no "vibes" in copy. - Handoff contract upgraded to **`drive_social_handoff: v2`**: `round`, `concepts_per_bucket`, `reserve_per_bucket`, `benchmarks`, `suggested_treatments`, `restricted_treatments`, and the `build` / `preserve` / `revise` / `correct` action lists (IDs always quoted — `"1.10"`). Removed - "Resolve everything here" as the operating principle; layout is now Step 3's decision. - The variation matrix, rotation rules, family census, and required format family / color mode / headline device / photo treatment fields. The vocabulary survives as an optional *Suggested execution* line marked provisional. - The "at least two no-photo concepts per bucket" quota and its "cheap insurance" rationale. - The `GEN-` plate brief table and the separate treatment-rotation table. - The fixed 10-concepts-per-bucket default. Changed - `brand_tile` is optional (with `brand_tile_status: approved | reference`); `people_policy` must be resolved before people-dependent concepts but is asked, not assumed. - v3 `permitted_families` / `permitted_color_modes` are read as suggested vocabulary only; genuine prohibitions inside them are moved into `restricted_treatments`. - Appendix A remains verbatim for traceability but is no longer the "source of record for the visual system" — Section 4 governs. - Tone rule: both registers by default, unless audience, subject, or an approved direction calls for one (say which). - `run_mode: iterate` is explicitly performance-driven scaling, distinct from replacement rounds. - v4.1 (same day): quoted YAML IDs, rejection scope, contextual tone rule, and `preserve` defined as a status-line change only — Codex's four compatibility refinements. - v4.2 (same day): the replacement round is explicitly **optional** — the default after a client review is to go live with approved ads and build fresh at the next refresh. Rejected ads are described in plain words (filename, headline, "the testimonial ones"); the skill matches them to concepts and confirms in a table. Every prompt states what it needs, why, and an example answer, for team members running the skill. **Client-provided assets are always approved for use, people in them or not** — `people_policy` governs AI-generated people only. - v4.3 (same day): all folder references made system-agnostic for Mac and Windows team members — the client folder is whatever the strategist points the skills at; no built-in default path. The team-distributed Codex skills carry the same change. ### `$brand-standards` — v3 → v4.0 (Codex) Added - Six-way evidence classification in the guide: client-approved requirement, strategist-approved direction, verified fact, **observed website styling**, proposed recommendation, unresolved question. The website is a source, not a ceiling for ad design. - Asset inventory with condition: usable-as-is, enhancement candidate, identity reference, or missing — with Keep real / Enhance / Generate / Must accomplish principles for Step 2. - `suggested_treatments` and `restricted_treatments` (sourced, scoped) in the v1 handoff; optional `brand/benchmarks/approved/` and `rejected/` folders. - `references/asset-intake.md` — Drive scan and download rules moved out of the main skill. Removed - Industry-to-font templates, fixed color area percentages, "exactly one accent" rules, family rotation, compulsory no-photo cards, and blanket bans on creative effects. - `permitted_families` / `permitted_color_modes` for new guides (legacy lists, if retained, are labeled suggested vocabulary). - Mandatory brand tile and the three brand-board prompts as required artifacts (both optional now). Changed - Rejected renders are recorded as specific execution failures, not universal bans. - The guide does not select production concepts, set counts, or emit action lists — Step 2 owns those. Handoff stays at v1; the skill version is separate. ### `$meta-ad-creative-generator` — v3 → v4.1 (Codex) Added - **Visual gate before presentation.** Every master is opened at full size and at ~360 px; five checks (idea and persuasion, composition, typography, brand and fidelity, finish and placement) each recorded pass / revise / blocked with observable evidence. Technical validity, creative review, and client approval are separate statuses. - **Execution brief per concept** written before rendering: visual idea, attention mechanism, image–message relationship, hero subject, composition, asset plan, and the failure to avoid. - **Handoff v2 routing** — `build` / `preserve` / `revise` / `correct` honored as disjoint lists; reserves and preserved concepts fail before any API call. New `scripts/handoff_contract.py` validates routing; `scripts/test_workflow.py` carries 33 offline regression tests. - **Protected originals** — `preserved-files.json` hash inventory checked before and after execution; approved files are never overwrite targets. - **Two production routes**, both subject to the visual gate: a fully generated master, or AI imagery finished deterministically with the original logo, licensed fonts, and editable layout tools (this was prohibited in v3). - **Reliable resume.** Each job writes a sidecar receipt with a fingerprint of prompt, ordered sources and their contents, model, quality, size, and candidate count; resume requires a matching receipt plus decoded dimensions and SHA-256. Changed jobs need new filenames or revision directories. - **v4.1 renderer:** default model is now `gpt-image-2.5-sunburst` with `quality: high` (`xhigh` / `max` available for specific unmet requirements). `gpt-image-2` and Flare remain explicit overrides, never silent fallbacks; switching models requires a new revision path. - Bounded exploration: up to three concepts × two candidates for a new direction, two internal repair rounds per concept, batches of at most five. Removed - **Stage A** whole-campaign background approval before any ad is built. - The single-ad smoke test as the sole gate; the 45-ad default target; the 32% band / 42% panel geometry; five-family minimum; big-type quota; asset-reuse cap; universal effect bans; the ban on local compositing and finishing. - Treating `gpt-image-2` output as the final master by definition. Changed - Similarity audit is advisory only — a duplicate warning, never a taste score or a redesign trigger. - Buckets' `asset_mode` is a summary; the per-concept asset plan is operative. - Rejected benchmarks establish specific failures, not universal bans; approved benchmarks guide craft without copying claims. --- ## v3 — 2026-09-10 · "Variety by layout" (superseded) - `/creative-strategy` v3: format families, color modes, headline devices, photo treatments, a per-bucket variation matrix, rotation rules, and two no-photo concepts per bucket — the response to "ten ads look like one ad." Fixed the symptom, not the cause; replaced by v4. - `$brand-standards` v3: brand tile, `permitted_families`, `permitted_color_modes`, `people_policy` added to the handoff. - `$meta-ad-creative-generator` v3: 4:5 masters at 1088×1360; per-job `quality`, `candidates`, `photo_required`; dry-run variety audit; new `scripts/similarity_audit.py`. ## v2 — 2026-09-02 (the version this site previously described) - Three-run pipeline with the single-file handoff, the bucket gate in Run 2, Stage A background approval and the one-ad smoke test in Run 3, 45 ads in batches of five.