SK Operating Playbooks Repeatable agent doctrine for J.A.R.V.I.S
Operating playbook layer

Operating playbooks turn agent talent into repeatable work.

The workbench gives the hands. Evidence gives the proof. Memory preserves context. Playbooks decide when to route, how to review, what to preserve, and where boundaries stay hard.

Open the routing deck Show the gates
16registered playbooks in the local inventory snapshot
11custom operator-created workflow lanes
SKILL.mdfrontmatter-driven trigger contract
0automatic durable mutation by default
Operating layer

The missing layer between tools and intelligence

Playbooks are not a list of prompts. They are the operating layer that makes Codex and J.A.R.V.I.S behave consistently across reviews, bug hunts, wiki work, Source of Truth updates, contracts, and handoffs.

workbench

The workbench

Search, read, edit, diff, shell, build, tags, task memory, and caller-bound verification surfaces.

Hands and controls
rag-core + wiki

The evidence and memory brain

Fetched corpus proof stays separate from durable wiki memory, with authority boundaries kept explicit.

Proof and recall
playbooks

The operating doctrine

Procedures, trigger routing, helper review, safety rules, and reusable behavioral patterns load at the moment of need.

Repeatable behavior
Trigger routing

One operator, many specialized lanes

The playbook catalog makes a broad agent feel like a team. Each request can land in a concrete workflow instead of relying on memory, vibes, or a one-off answer.

Review lane

The right reviewer wakes up first.

`review-helper` runs the local repo-grounded pass and routes bounded helper opinions through AGY, Codex, or ChatGPT handoff when useful. Findings stay hypotheses until locally verified.

review-helper agy-helper codex-helper chatgpt-helper
Custom operator inventory

The playbook wall already has serious workflows on it

The current custom catalog is focused on durable operator workflows, not one-off preferences. Each mature playbook gets wiki-backed recreation guidance so the behavior can survive beyond one machine.

kernelagents-md-compact-kernel-writerTurns AGENTS.md into a compact execution kernel with appendix detail.
helperagy-helperRead-only AGY/Gemini second opinion for reviews and triage.
huntbug-huntingParallel review, optimization discovery, and time-boxed find-and-fix rounds.
handoffchatgpt-helperBeads-backed human-mediated ChatGPT review requests.
helpercodex-helperFresh read-only Codex cold pass through an isolated wrapper.
pressure testgrill-with-docsPre-execution stress tests against repo evidence.
memoryrag-wiki-curatorSearch, read, classify, and curate durable internal wiki memory.
reviewreview-helperFindings-first broad review with bounded helper routing.
recoveryscratch-managerArchives and compacts oversized SCRATCH recovery logs.
truthsource-of-truth-curatorCreates, reviews, and updates project truth docs and ADRs.
contracttypescript-contract-guardPrevents prose-only API, schema, RPC, and wire-format contracts.
system supportsystem skillsImage generation, OpenAI docs, plugins, skill creation, and skill installation stay inventory-only unless needed.
Governance and safety

Behavior changes get treated like real engineering changes

A skill can change how every future agent acts. The skill-wiki-manager model keeps that power attractive without letting it become casual self-modification.

Skill second-pass review

High-impact behavior changes need a gate.

  • Trigger clarity and overlap are reviewed before install or wiki sync.
  • SKILL.md stays concise while scripts and references carry fragile detail.
  • Global-skill portability blocks repo-local leakage and machine-specific assumptions.
  • Helper output remains advisory until locally verified by the main operator.
Authority split

Playbooks do not own everything.

The active skill package owns behavior. The skills wiki owns durable explanation and recreation notes. The coding workbench may report readiness, but it must not mutate installed skills. Long-term memory stores context, not proof.

package authority wiki authority workbench boundary no auto-install
Compounding without runaway mutation

The harness can learn, but only as reviewable candidates

The accepted direction is not autonomous self-improvement. It is proposal-only distillation: useful session residue is classified, labeled, verified, and promoted only through explicit owner-approved work.

01

Observe

A session reveals repeated success, repeated failure, a durable decision, or a missing validation pattern.

02

Classify

The residue becomes a skill patch candidate, wiki memory candidate, SoT update candidate, follow-up task, or no action.

03

Label

Evidence, uncertainty, authority surface, safety boundary, and rejected paths stay visible.

04

Verify

Repo evidence, local checks, helper review, and normal gates decide whether the candidate is real.

05

Promote

Only explicit operator-approved work mutates skills, wiki, SoT, task state, git state, or runtime behavior.

Operating system layer

The agent does not just remember tools. It remembers how to operate.

Playbooks complete the operating stack: the coding workbench provides controlled action, the evidence engine provides grounded knowledge, and durable skill guidance turns repeated expert behavior into portable operating doctrine.