Document the Allium loop — gather context → take action → verify → repeat

[yavorpanayotov]

Allium, reframed as loop engineering

Boris Cherny (Anthropic, Head of Claude Code) has been blunt about where agentic coding is going: "I don't prompt Claude anymore. I have loops running that prompt Claude… My job is to write loops." And the single biggest quality lever in those loops is verification — he reckons a good verification feedback loop "2–3x's" the result.

That's exactly the gap Allium fills. This PR makes that explicit: it frames the five skills as the agentic loop — gather context → take action → verify → repeat — and positions Allium as the thing that makes the loop trustworthy: a behavioural verification signal, plus context that doesn't drift between sessions.

   gather context ───▶ take action ──────▶ verify ─────────▶ repeat
   /elicit · /distill   /propagate → tests   run tests          until
   → the spec           then implement       → /weed            converged
   (durable context)                         → CLI checks
          ▲                                        │
          └──────────────── revise intent ◀────────┘
              /tend, when verify shows the SPEC (not the code) was wrong

Why Allium is shaped for this:

• Gather context — the spec persists across sessions, so context doesn't drift (unlike re-reading code each time).
• Take action — /propagate projects the spec into tests, then you implement; for new behaviour, confirm tests fail first (red), so you're not writing already-covered or vacuous tests.
• Verify — run tests → /weed alignment → CLI structural checks: a behavioural pass/fail, not just "unit tests green."
• Repeat — convergence governs the whole loop, not just verify. When verification shows the spec was wrong, you revise intent (/tend), not only the code — an inner arc (code→contract) and an outer arc (contract→intent), mirroring how real development discovers requirements by building.

What's in here

• New reference skills/allium/references/recommended-loops.md — full walkthrough: the four-phase framing, the convergence invariant (two feedback arcs), Loop A (spec-first) and Loop B (code-first) diagrams, the red-step-before-implementation, autonomous-run guardrails (never weaken a test; escalate ambiguity), and the implementation prompt.
• README — a prominent ## The Allium loop section right after "How it works" (concept-forward, with background links).
• /allium router — a recommended-sequencing note so the assistant proactively suggests the next phase, and the reference added to the References list.
• Version bump — both plugin manifests 3.5.0 → 3.6.0 (minor; additive docs/guidance).

Validation

node scripts/test-skills.mjs → 54 passed, 0 failed (structure, links, routing, generation all green). generate-multi-editor.mjs --check clean.

Background: How Claude Code works · The New Stack — Loop Engineering · The Neuron — Claude Code creators on agent loops

🤖 Generated with Claude Code