Generate raster images and deliver them as verified files. The skill owns prompt craft, candidate discipline, artifact verification, ab-image-gen.result.v1 provenance and release QA on top of whatever image-capable model the harness can already dispatch; it does not own model access, credentials, billing or endpoint configuration. SKILL.md is the agent-behaviour contract; this README is repository-level orientation only, and no agent rule lives here alone.
| Field | Value |
|---|---|
| Skill id | ab-image-gen |
| Version | 2.0.0 (metadata.version in SKILL.md) |
| License | Internal |
| Status | 2.0.0 internal; no release tag and no CHANGELOG.md in the repo |
| Repository | https://github.com/archebase/ab-image-gen.git (branch main) |
- Owns: prompt compilation, bounded candidate sets, artifact save/verify,
ab-image-gen.result.v1provenance, deterministic crops, and the delivery verdict (ready/revise/blocked). - Delegates: the organisation's brand-guide skill, when one is installed — official logos, house typography and brand-guide rules stay outside this skill;
references/brand-production.mdsays to load that skill and let it own those rules. - Refuses: model choice, credentials, billing and automatic retries; official logos and final copy inside generated pixels; a second generation when a completed result already exists; any assumed model capability (edit, reference image, transparency, streaming) the harness has not demonstrated.
| Mode | Use when | Required evidence |
|---|---|---|
| Simple one-image request | A single deliverable; the full brief is skipped (SKILL.md, Workflow step 1) |
Image path, provenance path, model, actual format and dimensions, SHA-256 |
| Production work | A candidate set, crop family or consistency family; the full brief is used | templates/brief.md brief, generation record, templates/release-report.md verdict |
- Install the runtime:
python3 -m pip install -r requirements.txt(pinsPillow==12.0.0). - Read
references/generation-methodology.mdandreferences/prompt-compiler.mdbefore compiling the first prompt. - Ask the harness's image-capable model for the image — one distinct asset per request.
- Save the returned payload as a verified artifact; this writes the image plus its provenance sidecar.
- Inspect with
scripts/inspect_image.py, derive crops withscripts/make_variants.py, and reach forscripts/recover_harness_image.pyonly when the harness reports success but hands back no payload.
python3 scripts/save_image.py --input response.json --output out/hero.png \
--prompt "<the prompt used>" --model "<model id>"| Command | Purpose |
|---|---|
python3 scripts/save_image.py --input <payload> --output <path> [--prompt TEXT] [--model ID] [--provider LABEL] [--request-id ID] [--generation-id ID] [--overwrite] [--max-image-bytes N] |
Verify a payload, write the image and the provenance sidecar; refuses to overwrite an existing artifact or sidecar without --overwrite |
python3 scripts/recover_harness_image.py mark (--create-in DIR | --file PATH) |
Print or write the ISO-8601 UTC marker immediately before the harness call |
python3 scripts/recover_harness_image.py extract (--since ISO | --marker PATH) --out-dir DIR [--name STEM] [--harness codex] [--prompt-contains TEXT] [--wait-seconds F] [--all] [--keep-marker] [--force] [--provider LABEL] [--model ID] [--codex-home DIR] |
Find the completed call after the marker and save it through save_image.py; non-zero exit when no completed call is found |
python3 scripts/inspect_image.py <image> [<image> ...] |
Report format, dimensions, mode, byte size and SHA-256 as JSON |
python3 scripts/make_variants.py <master> --variant NAME=WIDTHxHEIGHT [--variant NAME=WIDTHxHEIGHT ...] --output-dir DIR |
Center-crop deterministic variants without generative changes; preserves alpha and refuses overwrite |
SKILL.md agent-behaviour contract: required path, rules, output contract
references/ 7 reference files on methodology, prompts, delivery, iteration, consistency, brand, briefing
scripts/ save_image.py, recover_harness_image.py, inspect_image.py, make_variants.py
templates/ brief.md, generation-record.md, release-report.md
tests/ test_save_image.py, test_make_variants.py, test_recover_harness_image.py
evals/evals.json 8 eval cases for this skill
.github/workflows/ validate.yml CI job
requirements.txt Pillow==12.0.0
SKILL.md owns the load triggers: generation-methodology.md and prompt-compiler.md before the first prompt, artifact-delivery.md before choosing an output path, iteration-and-review.md and consistency.md for iteration, and brand-production.md for an official logo, house typography or a brand guide.
python3 -m py_compile scripts/*.py tests/*.py
python3 -m json.tool evals/evals.json >/dev/null
python3 -m unittest discover -s tests -v
git diff --checkThese are the steps in .github/workflows/validate.yml. They show that the scripts and tests compile, that evals/evals.json is valid JSON, and that the three local test modules pass without network access or a credential. They cannot show model-output quality, live harness integration, or that a given commit's CI run succeeded.
git clone https://github.com/archebase/ab-image-gen.git
ln -s "$PWD/ab-image-gen" ~/.agents/skills/ab-image-genRequires Python 3.10+ and Pillow (requirements.txt pins Pillow==12.0.0). No credential or endpoint configuration is needed: the harness performs the model call, and scripts/*.py never make a network request and never read a credential. The repo has no release tags, so pin a commit (main head was dea64e0, 2026-10-01) if you need a reproducible checkout.
Current version is 2.0.0 (SKILL.md frontmatter); main head is dea64e0 "refactor(skill): rename image-gen to ab-image-gen". Verified in-repo: three test modules covering save/verify, variant cropping and recovery; evals/evals.json with 8 cases; the CI check list. Not verified here: no CI run result is recorded in the repo, no eval outcome is recorded, and the harness recovery path is exercised only against the codex adapter (ADAPTERS = ("codex",) in scripts/recover_harness_image.py).