From agent-skills
Spike lifecycle orchestrator — composes /launch-project, /write-doc, /validate-doc, /triage-feedback, /process-feedback, /revise-doc, and /lock-decisions to drive a full spike from research question to converged chapters with pending clusters surfaced. Use when starting a research spike or resuming one mid-iteration.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-skills:create-spikeThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
L4 lifecycle skill. You drive the full spike-creation flow by composing the L3
L4 lifecycle skill. You drive the full spike-creation flow by composing the L3 singletons and L2 dispatchers below. You do not author chapter prose, run reviewers, or apply edits yourself — every action is a delegated invocation of another skill.
rNN- files exist; chapters not converged)/create-plan/write-doc/create-decision (or /lock-decisions for
a finite list)/grill-meInvoke /launch-project. If it refuses (no project determinable from cwd or
prompt), ask the captain for a project name and invoke /launch-project once
more with that name. If the second attempt also refuses, abort with the
launch-project error verbatim — do not improvise a project context.
Once project context is loaded, identify the spike's research question and
proposed chapter list. If neither is in the captain's prompt, ask once before
proceeding. Use Bash to scaffold via
${CLAUDE_PLUGIN_ROOT}/scripts/yolo new spike <project> <name> — never write the
spike README.md path directly. Parse the JSON envelope's filePath to
discover the assigned sNN-<name>/README.md. Per-chapter authoring uses
${CLAUDE_PLUGIN_ROOT}/scripts/yolo new chapter <project> <spike-slug> <chapter-name>,
which returns the assigned NN-<chapter-name>.md path the same way.
For each proposed chapter, in declared order:
Invoke /write-doc with the chapter file path and a directive carrying the
chapter's purpose, sources, and required sections. /write-doc produces the
file; you do not write prose yourself.
Invoke /validate-doc with:
artifact path = the chapter filefocus list = [accuracy, architecture, depth, dx] (the create-spike
baseline per the validate-doc focus-list table). Add security for
trust-boundary chapters; extend per chapter nature; never trim below
the baseline four.iteration number = current iteration for this chapter (start at 1)review directory = the chapter's parent directoryBlock until validate-doc returns its summary. Surface any errored reviewer
verbatim — do not auto-retry, do not silently proceed without it.
Invoke /triage-feedback with the rNN-{focus}.md paths validate-doc
landed and the chapter's path. Receive the cluster list with per-cluster
routes (act / auto-act / process / defer / lock-decisions).
Walk the cluster list and dispatch per route:
act / auto-act — invoke /revise-doc with the chapter path and
the rNN- paths whose findings drive the cluster. revise-doc handles the
edits; you record the result.process — dispatch /process-feedback subagents in parallel,
one per process-routed cluster, all in a single Task batch (cap
4 per batch; if more, batch sequentially in rounds). Each subagent
returns either verified-resolution or narrowed-options-brief. For
verified-resolutions: invoke /revise-doc with the patch suggestion
applied as the act-pass. For narrowed-options-briefs: surface the brief
to the captain inline if the cluster is blocking; carry forward as
non-blocking otherwise.lock-decisions — collect every cluster routed to lock-decisions
in this iteration into a single findings list, then invoke
/lock-decisions once with that list. Record per-finding outcomes
(locked / deferred / grilled-then-locked / grilled-then-deferred).defer — record the cluster as deferred. No action.The chapter is converged for this iteration when every routed cluster has
landed (act/auto-act applied; process returned a verdict and either
auto-applied or surfaced; lock-decisions returned outcomes; defer
recorded).
Iterate the chapter — re-author touch-ups via /write-doc if needed, then
back to step 2 (Validate). Convergence rules:
A chapter exits the loop when an iteration's validate-doc lands no new blocking clusters AND every prior cluster is resolved or in pending state.
After all chapters converge per-chapter, run an overall pass:
/validate-doc with artifact path = the spike README.md,
same baseline focus list, fresh iteration number./triage-feedback on the rNN- files validate-doc landed at
the README level.When all chapters and the overall pass converge, present a markdown summary to the captain with these required sections:
/lock-decisions invocation: cluster ID,
topic, outcome, dNN- path (if locked)acknowledged-without-decision / blocked-by-X /
awaiting captain). Empty list explicitly stated as "No pending
clusters."A "converged" report that hides pending clusters violates the convergence-summary requirement and is forbidden.
process-feedback dispatch — process-routed clusters
must dispatch in a single Task batch (cap 4); sequential dispatch
defeats the parallelism budget and is a known eval failure./launch-project when invoked from outside a project
context — the pre-flight call is non-negotiable./write-doc
— this skill is an orchestrator; it does not write prose./revise-doc reports a contradiction, surface it to the captain;
do not pick a winner./grill-me's job, invoked via /lock-decisions on
the captain's grill choice.npx claudepluginhub mikestopcontinues/agent-skills --plugin agent-skillsCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.