Distillery

The context brain for a fleet of unattended coding agents. This repo is the source of truth for the context files, directives, and skills that drive igor (the harness) and the operator's local harnesses -- extracted out of the harness so the same brain can pour into more than one glass. Context is treated as first-class, versioned, validated artifacts that agents consume, rather than prose that rots inside whatever repo it started in.

Mirrors

Canonical repo: https://git.sherver.org/joshtronic/distillery

Mirrored for your convenience:

• https://github.com/joshtronic/distillery

I don't monitor these services or accept pull/merge requests on them.

How it works

• skills/<name>/SKILL.md -- one distillable unit: YAML frontmatter (name, description, consumers, source, extracted) followed by the content. Extra files ride in skills/<name>/files/.
• bin/still -- the assembler. still validate enforces the format contract deterministically; still verify runs each skill's optional verify: command as a semantic gate on top of that -- VERIFIED, FAILED, or honestly UNVERIFIED when a skill has no check yet; still build refuses to ship anything unvalidated and writes dist/manifest.json; still list shows the shelf.
• dist/manifest.json -- a committed, generated artifact (the rest of dist/ stays gitignored). A consumer verifies each poured skill's sha256 against this file read at a ref, so it must exist on that ref -- regenerating it from the checkout being verified would prove nothing. Regenerate it with bin/still build whenever a skill changes, and commit the result; bin/still build --check fails CI if the committed copy has drifted from the tree.
• Releases -- not cut yet. Consumers read commits directly off origin/master, never a working tree.

The shelf

Generated from each skill's description frontmatter by bin/still index; still index --check fails CI if this block drifts from the tree. Run bin/still list for the same listing from your terminal.

• ai-writing-tells — Catalog of AI writing tells to avoid in any prose surface
• coding-standards — Use when writing or reviewing any code change. The coding-output contract -- minimal touch, the two-variety comment rule, TDD with red-tested validation, honest verification claims.
• design — Use when doing visual or UX work on a site -- new UI, redesigns, reskins, or design review. Distinctive-per-subject direction with content preservation and verification-by-screenshot.
• doctrine — The operator's hard-won engineering doctrine for agentic systems -- amnesia, deterministic gates, mechanism over prose, red-tested validation, pre-mortems, evidence over claims, guard severance
• dossier-spec — The dossier's declared-shape vocabulary -- AGENTS.md structure, the Metadata key contract (url, automerge.require_human, automerge.maintenance, landed.kind), and the two invariants (explicit opt-in with fail-fast on partial declarations; no unattended edits to the dossier itself)
• feedback-directive — Player-feedback triage directive: DROP/FILE decision contract over untrusted CSV rows
• now-directive — Weekly /now page refresh directive
• product-research — Use when delegated a product goal for a site or tool -- researching whether and how to pursue it, and producing specced, verifiable tickets. Replaces the retired autonomous-CEO pattern with demand-driven product work.
• review-directive — The shadow code reviewer's directive: verdict contract, review dimensions, diff-appropriateness, comment-bloat findings
• site-work-directive — Weekly website work-pass directive
• sports-digest-directive — Daily sports digest directive: teams-first curation with league news as fallback, taught-concepts curriculum
• ticket-skeleton — The fixed skeleton every task spec (issue / ticket) is written against -- goal, out-of-scope, verification, done-means; the amnesiac worker reads nothing else
• voice — Igor's shared voice anchor for all writing surfaces
• worker-contract — The unattended worker's operating contract: outcomes (PR/report/block), PR_BODY discipline, TDD, security self-review, scope rules
• worker-permissions — The unattended worker's tool permission profile (Claude Code settings) -- broad dev capability, denied network git / secrets / harness internals

Consuming

No release is cut yet -- consumers currently read origin/master directly. Once one exists, read dist/manifest.json at that ref and take the skills whose consumers list matches your surface (igor-surface for harness prompt fragments, generic for anything). Wiring igor to consume a release instead of its own copies is tracked on igor's side.

Testing

make test

Runs bin/test-still.sh, which exercises every validation rule by deliberately breaking it -- a validator that has never failed has never been tested -- and finishes by validating the real tree.

License

GPL v3 -- see LICENSE.