Operationalize

/operationalize — Distill + Route Bridge

Loop position: move 7 (capture + ratchet) of the operating loop — routes promoted learnings to their weakest durable enforcement (skill, gate, or bead).

Rich context dies in the artifact that gathered it. A deep-research report, a codebase-recon sweep, or a painful learning is read once, agreed with, and never changes behavior again. This skill is the bridge: distill the artifact into a handful of evidence-anchored rules, then route each rule to the automation shape that will actually fire next time — skill, workflow, hook, gate, beads, or playbook.

Use when: "I gathered rich context — operationalize it." The input is a finished artifact; the output is rules with anchors and a handoff per rule.

⚠️ Critical Constraints

• Sources stay in place — to prevent a corpus-curation detour. Name the

source artifacts by path; never copy them into a corpus directory. The artifact you were handed IS the evidence base.

• Every rule cites a source anchor, because an unanchored rule is an

opinion wearing a rule's clothes — it cannot be audited, challenged, or retired when the source is superseded.

• Disagreement is marked DISPUTED, never averaged, because splitting the

difference between conflicting sources produces a rule nobody measured. A DISPUTED entry routes to investigation (beads), not to automation.

• Shape is decided by routing, not by vibe — to prevent everything

becoming a skill. Compose /automation-shape-routing for the shape decision; this skill only extends its target list.

• Gates start warn-only, because a fresh rule promoted straight to a

blocking gate ships its false positives as outages. Promotion to blocking comes after the gate has run quietly on real traffic.

• No rule survives without a counter-example check, because the cheapest

time to find the case where the rule is wrong is before it is wired into anything (see Step 5).

Execution

Step 1: Intake

Name the source artifacts in place — absolute or repo-relative paths plus a one-line provenance note each (who produced it, when, method). Confirm each source has citable anchors (section IDs, finding IDs, line ranges); if not, add anchor IDs to your notes about the source, never by editing the source.

Checkpoint: every source is a named path with a provenance line. No corpus dirs were created.

Step 2: Distill

Extract candidate rules in the canonical form — "When X, do Y because Z" — where Z cites at least one anchor. Work source by source, then reconcile:

• Multiple sources agree → one rule citing all supporting anchors.
• Sources conflict → one DISPUTED entry naming both sides' anchors and

what evidence would settle it. Do not synthesize a compromise rule.

• A finding with no behavioral consequence → drop it (context, not a rule).

Checkpoint: every rule line carries ≥1 anchor; every conflict became a DISPUTED entry, not a blended rule.

Step 3: Route

Hand each rule to /automation-shape-routing and extend its decision with this target table:

Checkpoint: every rule has exactly one route; every DISPUTED entry routed to beads.

Step 4: Emit

Write the rule packet (Output Specification below), then create one handoff stub per routed rule: the rule text, its anchors, the chosen route, and the target skill invocation. The downstream builder owns the artifact; this skill owns the rule and its evidence trail.

Step 5: Validate

For each rule, run the counter-example check: actively search the sources (and your own experience) for one case where following the rule would be wrong. A found counter-example narrows the rule's "When X" or demotes it to DISPUTED. Then request a /validate verdict on the packet before handing off — verify before any downstream builder consumes it.

Worked example (golden fixture)

Input: fixtures/research-excerpt.md — a fake deep-research excerpt on worker-lane retry behavior, anchors RX-1…RX-5.

Distilled packet:

1. When a lane receives a rate-limit response, wait the full advertised cooldown before any same-account retry, because burst retries extend the penalty window (RX-1) and sub-30-second retries re-failed in 84% of logged events (RX-2). → route: hook (mechanical, event-triggered). 2. When one same-account retry has already failed, rotate accounts before the next attempt — capped at three accounts per hour, because post-failure rotation succeeded on the next call in 91% of cases (RX-2) while sustained cycling beyond three accounts/hour risks account review (RX-5). → route: skill (judgment about when the cap binds). 3. DISPUTED — rotate unconditionally on the first rate-limit response. RX-4 (operator interview) asserts it; RX-5 flags rotation frequency itself as a risk, and RX-2 only measured rotation after a failed retry. Settling evidence: telemetry comparing first-response rotation vs post-failure rotation. → route: beads (investigation), not automation.

Note what did NOT happen: rules 1–3 were not averaged into "rotate fairly quickly"; RX-3 (re-dispatch to a warm lane) was held back at Step 5 because its own source records a 9% duplicate-work counter-example.

Output Specification

Format: markdown rule packet — sources-in-place list, numbered rules in "When X, do Y because Z" form with anchors, DISPUTED section, route table, and the validate verdict reference. Path: written to .agents/operationalize/YYYY-MM-DD-<slug>.md; handoff stubs accompany it as a ## Handoffs section (one block per routed rule). Exit signal: packet path + per-rule route summary reported to the caller.

Quality Rubric

• [ ] Every rule cites at least one stable source anchor
• [ ] Zero blended rules: every source conflict appears under DISPUTED
• [ ] Every rule has exactly one route, chosen via automation-shape-routing
• [ ] Any gate route is explicitly marked warn-only
• [ ] Counter-example check ran per rule and is recorded in the packet
• [ ] No source artifact was copied or moved; no corpus directory exists

See Also

• automation-shape-routing — the shape decision this skill composes
• skill-builder, workflow-builder, cc-hooks, beads-workflow — emit targets
• validate — the packet verdict before handoff
• research — typical upstream producer of the input artifact

Research excerpt — worker-lane retry behavior under provider rate limits

Golden fixture for /operationalize. Fake but representative deep-research

output: three sources, stable anchors, one genuine disagreement.

Findings

[RX-1] (source: provider-docs, 2026-05) Burst retries against a 429 extend the penalty window. Both surveyed providers document that immediate re-send after a rate-limit response resets the cooldown clock; clients that waited the full retry-after header value cleared in one cycle.

[RX-2] (source: swarm-telemetry, 2026-06) Across 312 logged rate-limit events in overnight swarm runs, lanes that retried inside 30 seconds hit a second 429 in 84% of cases. Lanes that rotated to a different account first succeeded on the next call in 91% of cases.

[RX-3] (source: swarm-telemetry, 2026-06) Re-dispatching the in-flight work item to a different, already-warm lane resolved 61% of rate-limit stalls faster than waiting out the cooldown on the original lane — but produced duplicate work in the 9% of cases where the original lane later resumed on its own.

[RX-4] (source: operator-interview, 2026-06) The operator reports that account rotation "always beats waiting" and recommends rotating on the first 429, unconditionally.

[RX-5] (source: provider-docs, 2026-05) One provider's terms flag rapid credential cycling as abuse-signal behavior; sustained rotation across more than three accounts in an hour risks a account-level review.

Noted disagreement

RX-4 (rotate unconditionally, immediately) conflicts with RX-5 (rotation frequency itself carries account risk). Telemetry (RX-2) supports rotation but measured it only after a first failed retry, not as a first response.