First Steps — env
Add the env profile on top of core: tier criteria on disk and one idea taken through /arch — six steps rendered from the First-Steps source of truth.
Env is core plus the multi-model contour: the tier-routing criteria land on disk as a
file you can read, and you take one idea through /arch to see the design contour work.
Experimental. The sequence's goal, from the source: core + the tier criteria on disk
and one idea through /arch.
Run it with the profile flag --profile env.
The six steps
- Install at env depth —
bash <getff>/install.sh <stack> --profile env. Already on core? Re-run that same command: the deeper payload is added and every core artefact stays byte-identical,.prettierignoreincluded — its managed block ships full-depth from core time. A bare--refreshnever deepens — with no--profile, a refresh resolves to core depth, re-delivers fixes to what is already on disk and creates nothing new. To refresh and deepen in one pass, ask for it explicitly:--refresh --profile envdeliverstier-home.mdand the env skills.
- Verify the payload landed —
ls AGENTS.md .ai-factory/ scripts/, plus.ai-factory/tier-home.mdand the five skillsenvadds overcore:.claude/skills/arch/,night-mode/,orchestrator/,pipeline/andreviewer/. Those are the artefacts the rest of this sequence uses.
- Fill the project passport — replace every
<…>placeholder field in.ai-factory/DESCRIPTION.md. If you installed straight toenvthis is the same passport step as core's; if you deepened an existing install, it is worth re-reading what you wrote there before the design contour starts reading it.
- Prove the rules are not inert on your layout —
bash scripts/check-rule-globs.sh, thenbash scripts/check-fences-fire.shto see a rule go RED on planted input. On an empty skeletoncheck-rule-globs.shfails by design (zero source files to match); re-run it once you have some. Same gates as core, same reason: a rule nobody has seen fire is an unproven claim.
- Read the tier + degradation SSOT — open
.ai-factory/tier-home.md. It owns the Tier 0/1/2 routing criteria and the capability-absence degradation matrix: what still works with no aif runtime, no executor-tier model, or a non-Claude-Code harness.AGENTS.mdonly points there; the criteria live in that one file. The degradation half is rendered publicly on our Degradations page.
- Take one idea through
/arch— invoke/arch <topic>: the external design contour turns a raw idea into a reviewed design plus a routed handoff. On a harness without skills, read.claude/skills/arch/SKILL.mdand follow it by hand. Pick a small real idea from your own backlog — the point of the step is to see the contour produce a design you would actually merge.
Where this goes next
What the tiers route to, one level up: Overview — multi-model pipeline. Or continue the sequence at First Steps — factory, which drives one task through the pipeline end to end.
First Steps — core
Install getff at core depth and end with a rule that has gone red on planted input in your own repo — seven steps rendered from the First-Steps source of truth.
First Steps — factory
The factory profile: one task driven through the dispatch pipeline end to end — eight steps rendered from the First-Steps source of truth.