Menu
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/nirecom/agents --skill workflow-initThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Edit source code for the current task. Delegates editing and lint/typecheck/self-repair to a subagent.
Plan and write test cases with high reasoning effort. Test iteration runs in a subagent to minimize confirmations.
Conduct a decision-tree interview with the user to lock in requirements, motivation, scope, and non-goals before planning.
Showing SKILL.md
| name | workflow-init |
| description | Routing entry point for the Claude Code workflow. Inspects |
| model | sonnet |
IMPORTANT: Interactive session required. Hard-fail in non-interactive contexts (claude -p, /loop, subagents).
First step of every workflow session. Routes on GH issue context: Path A (#N + intent:clarified) skips interview; Path B (#N, no label) pre-fills interview; Path C (no #N) runs full interview and auto-creates a tracking issue.
Read rules/github-issues.md before WI-2 — on-demand-only, never auto-injected; the meta label, intent:clarified, closes_issues insertion order, and the sub-issue open-state gate all come from it.
Canonical: skills/_shared/resolve-plans-dir.md. Substitute the resolved absolute path for every <PLANS_DIR> placeholder below. Subagent prompts must receive literals — they cannot expand $VAR.
The driver (bin/workflow/workflow-init-driver) handles WI-3..WI-9: token detection, gh issue view fetch for each N in ISSUES, CLOSED detection, label extraction, meta classification (meta strip / open sub-issue guard), Aggregate WIP check (all N: all_same / all_none / any_other), route decision, and context.md write. The driver writes the checkpoint and <SID>-context.md directly under PLANS_DIR (outside git repos → ENFORCE_WORKTREE does not apply).
Invocation: node "$AGENTS_CONFIG_DIR/bin/workflow/workflow-init-driver" <raw-tokens> 2>/dev/null
(<raw-tokens> = issue number tokens like #15 or #15 #22 — not the full user prompt)
On resume (after ask_user): node "$AGENTS_CONFIG_DIR/bin/workflow/workflow-init-driver" --resume <CHECKPOINT> --answer '<token>'
Read all KEY=VALUE output lines. Dispatch on ACTION=:
| ACTION | Meaning | What to do |
|---|---|---|
invoke | Next skill ready | Run NEXT_SKILL via Skill tool |
done | Driver complete | Proceed to WI-10; follow PATH_DECISION= at WI-12 (Path C still runs WI-10 onward — no jump) |
blocked | Unrecoverable | Show REASON= + NEXT_HINT= and stop |
ask_user | User decision needed | Percent-decode OPTIONS_DISPLAY= (encoded like QUESTION=), present via AskUserQuestion; re-invoke with --resume <CHECKPOINT> --answer '<token>' |
emit_sentinel | Driver requests sentinel | Emit SENTINEL= via separate Bash call |
ask_user interruptions and their options:
ASK_ID=wip_conflict: Issue(s) # are in progress in another session. Driver answers: Continue (recommended) / Abort. On Continue: for each N in ISSUES, driver runs wip-state set <N> (override for other; claim for none; idempotent for same). On Abort: ACTION=blocked REASON=user_aborted.ASK_ID=wip_rc2: wip-state set rc=2 for #N. Driver answers: Continue (acknowledge risk — driver treats as none, proceeds) / Abort.ASK_ID=wip_error: wip-state check failed (transient auth / session-id resolution failure — check $CLAUDE_ENV_FILE or $CLAUDE_SESSION_ID, rc=non-zero). Driver answers: Continue (treat as none, proceed) / Abort. Advisory: driver logs [wip-state check failed for #N — proceeding as 'none'].ASK_ID=closed_reopen_<N>: Issue #N is CLOSED. Driver answers: reopen / remove / abort. Remove is only offered when len(closes_issues) >= 2.ASK_ID=meta_select: meta issue has open sub-issues. Driver answers: #<M> (select sub-issue) / abort.ASK_ID=fetch_failed_path_c: gh issue view fetch failed. Driver answers: continue (Path C) / abort.PATH_DECISION= values on ACTION=done:
A — all N carry intent:clarified (and FORCE_PATH_B is not set from a fresh WIP claim).B — any N lacks intent:clarified OR WIP was freshly claimed for all N (FORCE_PATH_B=true). ALL_NONE path: none + for each N in ISSUES, wip-state set <N> (from fresh claim), then set FORCE_PATH_B=1 — even when intent:clarified is present, FORCE_PATH_B routes to B. ISSUES[0] becomes closes_issues[0]; all entries become closes_issues in insertion order. ISSUES[@] are processed symmetrically.C — zero issues (ISSUES=()) OR NON_GITHUB=1.META — all issues carry meta label and have no open sub-issues.Dispatch the survey-code + survey-history subagents per skills/_shared/subagent-concurrency.md SC-P (independent — distinct artifact paths, neither reads the other's output). For each: prompt session-id=<resolved>, context_path=<PLANS_DIR>/<session-id>-context.md, artifact_path=<PLANS_DIR>/<session-id>-survey-{code|history}.md, instruct to read context_path + follow skills/<survey-code|survey-history>/SKILL.md Procedure, write to artifact_path, do NOT invoke make-outline-plan. Substitute <PLANS_DIR> with the absolute path resolved at WI-1.
Note: the Stop-guard silence during dispatch is automatic (PostToolUse marks the step in_progress). Do not emit NEXT_STEP_PAUSE.
Apply skills/_shared/survey-artifact-valid.md to each artifact. On invalid: emit <<WORKFLOW_SURVEY_AGENT_FAILED: survey-code>> or <<WORKFLOW_SURVEY_AGENT_FAILED: survey-history>>. Fall through to WI-12 on failure — do NOT abort. clarify-intent handles missing-or-invalid artifacts.
WI-8 open sub-issue guard ensures all ISSUES[@] have no open sub-issues before reaching this path.
node "$AGENTS_CONFIG_DIR/bin/workflow/set-workflow-type" --session "$SESSION_ID" --type wf-meta --advance --step workflow_init --complete (single Bash call — records workflow_type and completes workflow_init together; no --next since the next action is fixed at PM2).echo "<<WORKFLOW_CLARIFY_INTENT_NOT_NEEDED: meta issue — WF-META type; intent confirmed from issue body>>"./issue-create --skip-survey with --verdict bulk-sub-of --parent <meta-N> --manifest <file> to create all planned sub-issues under the meta parent in a single bulk pass.make-outline-plan. (next-step auto-skips detail and 8 other non-applicable WF-CODE steps after outline completes — make-detail-plan is never invoked in WF-META.)<PLANS_DIR>/<session-id>-intent.md (strip sentinels from body): # Agreed Requirements — <session-id>, then immediately a **Title:** <human-readable one-line title> line (same data contract as clarify-intent CI-4), ## Issues (one - #<N>: <title> line per entry in ISSUES[@], in insertion order, no annotations), ## Background / Motivation, ## Scope / Constraints, ## Accepted Tradeoffs (none — capture at outline stage). Title for each N from WI-4's gh issue view; fetch failure → - #<N>: (title unavailable). Never omit ## Issues or ## Accepted Tradeoffs — latter is detail-planner.md Approved Scope gate. ## Issues is SSOT for closes_issues (canonical parser: hooks/lib/parse-closes-issues.js). ISSUES[0] is the first entry; it becomes closes_issues[0]; this entry becomes closes_issues[0].closes_issues into session state: node "$AGENTS_CONFIG_DIR/bin/parse-closes-issues" --session "$SESSION_ID" (separate Bash call; routes through the write-once cache in hooks/workflow-state/session-facts.js).node "$AGENTS_CONFIG_DIR/bin/cc-session-title" set-issue "$(pwd)" "<PLANS_DIR>"skills/workflow-init/scripts/path-a-label-and-board.sh with "${REPO_MAP_ARGS[@]}" followed by all entries of ISSUES[@] as positional args; export PLANS_DIR, SESSION_ID, AGENTS_CONFIG_DIR. Adds intent:clarified (--add-label "intent:clarified") to each related entry (fail-closed — on failure writes ABORT marker <PLANS_DIR>/<session-id>-workflow-init-aborted-pathA-multiN-label-failure.md + exit 1). For every issue it runs ensure-board-card.sh (best-effort, warn-and-continue). Both idempotent.node "$AGENTS_CONFIG_DIR/bin/workflow/next-step" --advance --step workflow_init --complete (no --next — the next action is fixed at A3a, not next-step's judgment), then, unchanged, as a separate Bash call: .When
--advanceis called without--next, noACTION=block is ever emitted — proceed directly to the next documented step above rather than waiting onnext-step's judgment.
<PLANS_DIR>/<session-id>-issue-prefill.md with <!-- Issue #<N> seed for clarify-intent. Confirm framing, do not start from scratch. -->, # Issue #<N>: <title>, <body>.echo "<<WORKFLOW_MARK_STEP_workflow_init_complete>>" (separate Bash call).clarify-intent with #<N> in args so step 1a auto-detect fires.echo "<<WORKFLOW_MARK_STEP_workflow_init_complete>>" (separate Bash call).clarify-intent (no pre-fill).Cannot be skipped. For docs-only bypass: echo "<<WORKFLOW_MARK_STEP_workflow_init_complete>>" directly.
echo "<<WORKFLOW_CLARIFY_INTENT_NOT_NEEDED: issue #{N} has intent:clarified label>>"skills/_shared/judge-task-complexity.md; evaluate all S1–S6 signals against the confirmed intent.md (S6 approximated from intent.md line count only — outline.md does not exist yet). Also evaluate so_c1/so_c2 (criteria: skills/_shared/judge-plan-skip.md) from intent.md — needed regardless of the internal SKIP_MODE, since that is not known until the call returns. Run as a single Bash call: bash "$AGENTS_CONFIG_DIR/bin/workflow/record-complexity-and-skip" --session "$SESSION_ID" --verdict <high|low> --signals <csv-or-empty> --target outline --advance --so-c1 <true|false> --so-c2 <true|false>. Read the SKIP_DISPATCH=<value> line straight from that call's own stdout for the next step's dispatch — do not capture it into a shell variable ($(...)) or pipe it (|): the Bash tool resets shell state between calls, so a captured variable would not survive into that separate call anyway. SKIP_DISPATCH is no-skip / advanced / need-judgment — the wrapper records the complexity evaluation and, when its internal SKIP_MODE=auto, also settles outline as skipped via --advance in the same call.SKIP_DISPATCH=no-skip → A4 (outline not skipped). SKIP_DISPATCH=advanced → outline already settled skipped by A3a's --advance call; dispatch Agent tool (run_in_background: true): subagent_type=skip-verifier, session_id=$SESSION_ID, target=outline, intent_path=<PLANS_DIR>/$SESSION_ID-intent.md → A4. SKIP_DISPATCH=need-judgment (internal SKIP_MODE=judgment with so_c1/so_c2 both true) → run node "$AGENTS_CONFIG_DIR/bin/workflow/record-skip-judgment" --session "$SESSION_ID" --target outline --advance --c1 true --c2 true as a separate Bash call (no --next; settles outline as skipped) → dispatch the same Agent tool call → A4.make-outline-plan (surveys already complete via WI-9).