foundry

Foundry

This repo converges 84 upstream design-engineering skills into six (pe-design, pe-build, pe-review, pe-verify, pe-product-description, pe-brand-assets), preserving the best source prose byte-for-byte and proving it with hashes. You are operating a factory with strict rules, not editing a normal repo. AGENTS.md is the fence; this skill is the manual.

Before anything else

1. Read foundry/MANIFEST.json — look up the class of every file you intend to touch.
2. Read foundry/LEDGER.md — rulings bind everything you author.
3. Pick the runbook that matches the task (table below) and follow it exactly.
4. End every session that changed provenance state with a foundry/LOG.md line and a green foundry/scripts/check-integrity.sh.

File classes

Class	Meaning	Your permissions
verbatim	Byte-for-byte lift from a pinned upstream commit	Never edit. Re-lift via script or nothing.
verbatim-minus	Verbatim with a recorded cut (pristine + .patch in foundry/pristine/)	Never edit the file directly. Changing the cut = edit file, run make-patch.sh, PR shows the patch diff.
distilled	Authored condensation of recorded sources; the salvage list in its derivation file is the contract	Edit only via the distill-refresh runbook; finish with check-integrity.sh --update-distilled <path>.
authored	Ours: spines, glue, foundry docs	Edit freely, subject to the ledger and the anti-slop gates.

private: true entries are course drop-in files — purchased content that exists only on local machines, never in a commit.

The convergence method (for building or revising a skill)

• Architecture: each skill is a thin authored spine (router: modes, cues, contract, rulings block; target < 150 lines) over references lifted verbatim from graded sources. New prose is written only for glue, routing, and rulings.
• Grades bind handling: GOLD → lift verbatim, never reword (cut whole sections only, via verbatim-minus, logged) · SOLID → lift, dedupe only · PADDED → distill against a named salvage list · SLOP → exclude.
• Anti-slop gates on anything you author: no restatement between spine and references (principle in the spine, recipe in the ref, never both) · every sentence must change an implementation choice · no adjective stacks, personas, or dial theater · skill activation load stays under ~500 lines · no provenance in runtime text — MANIFEST, derivations, and NOTICE own provenance; skill prose states rules without citing sources, patches, pins, or ledger ids (frontmatter carries only a one-line derivations pointer).
• Plan first: write foundry/derivations/<skill>.md (sources, classes, cuts, rulings applied) before running a single lift. The receipts precede the work.
• Verify: check-integrity.sh green, skills-ref validate (or equivalent) clean, spine checked against every ledger ruling it touches.

Runbooks

Task	Runbook
Bring a file in from an upstream	runbooks/lift.md
An upstream changed a vendored file	runbooks/re-lift.md
A distilled file's source changed, or the distillation needs work	runbooks/distill-refresh.md
Sources contradict, or a ruling needs to change	runbooks/new-ruling.md
CI integrity check is red	runbooks/integrity-failure.md
Install/refresh the purchased course files locally	runbooks/course-dropin.md
Evaluate a new upstream or add a skill to the kit	runbooks/new-skill.md

Division of labor — never blur it

LLMs propose (grade, classify drift, author spines, draft PRs) · hashes verify (check-integrity.sh, in CI on every push) · humans merge (nothing you produce lands on main without review). Your token stream is never the channel vendored prose travels through.