requirements-inline-skill
In-session requirements delegate — runs the AC-quality detection gate, grills reviewer-emitted Requirements Gaps inline (Mode R, agent file as checklist), and drafts BRD/SRS in-session with native que
SKILL.mdskills/requirements-inline-skill/SKILL.md
GitHubRequirements (inline)
You are executing the requirements specialist's workflow in this session. The subagent's isolation and its headless relay dance (Modes A/B exist precisely because a spawned session cannot ask questions) are replaced by this session's native questioning — apply the discipline below, and never delegate the work.
You are a wrapper, not a copy. Mode conduct lives in the deployed
agents/requirements-specialist-subagent.md (loaded as checklist); BRD/SRS
templates live in civiltekk-requirements-specs-skill; interrogation
methodology lives in grilling-skill. This file owns the detection decision
tree and the invocation loop only.
Detection decision tree (pipeline route)
Inputs: the ticket's acceptance criteria, the PLAN's AC section, the Step 7 triage outcome (reviewers-selected count), and whether the ticket is new work or a retry.
Skip rule (default): AC quality passes AND at least one reviewer was selected → report
skip(one line: why), continue to reviewer triage.Branch hinges on judgment (thin-vs-substantive ACs, ambiguity calls)? Optionally consult the
laya_classifytool with the statement — it returns calibrated p(yes) per question (local, offline); weigh the probabilities and decide. Deterministic checks (field presence, counts, template shape) stay in code — never classify what a rule already answers.AC-quality rule: run the interrogation route when ANY holds — fewer than 3 ACs; ambiguity markers (vague quantifiers like "fast"/"most", "and/or" coupling, no verifiable predicate an implementation can pass/fail); or two ACs that conflict.
Zero-reviewer rule: triage selected no reviewers AND the ticket is new work → run the interrogation route regardless (there is no rider duty to carry requirements verification).
Interrogation route: apply
grilling-skillmethodology to the suspect ACs — one question at a time through the session's question capability, each with a recommended answer grounded in the ticket body and PLAN. Apply answers to the ticket body and/or PLAN ACs (edits land in the caller's working tree; list every file changed). Cap: 2 rounds, then surface the residue and proceed on documented defaults.
This gate is data-driven AC interrogation before reviewers run. It is NOT a PLAN review by requirements-specialist (the stage-mismatch decision stands — Mode R's designed input remains reviewer-surfaced gaps).
Mode R route (review-gap relay)
Resolve the checklist (§Checklist resolution), load reviewer-baseline-skill
conduct, then per the checklist's Mode R: for each gap, sharpen
suggested_question into one concrete question with a recommended answer
grounded in the gap's evidence. Return Questions for the user — one per
gap. Max 2 relay rounds per review (caller-enforced; state the round number).
Mode A route (in-session drafting)
Resolve the checklist, then per its workflow (Steps 0–5) with the interactive
caveat inverted: you HAVE the question channel — use it instead of relay
rounds. Doc-type routing tree, templates, and file conventions come from
civiltekk-requirements-specs-skill (matching brd/srs route). Drafts
land at docs/{brd|srs}/{BRD|SRS}-draft-{slug}.md in the caller's working
tree.
No-subagent pin
This skill runs fully in-session. Do NOT delegate requirements work to a
subagent — no Task calls, no child sessions, no Mode A/B relay round-trips.
The isolated-child variant is a different invocation
(requirements-specialist-subagent via the v1 arm or direct delegation),
not an escalation path available here.
Checklist resolution
Capability binding (per the portability contract — the agent self-selects its row; unknown harnesses fall through to the fallback):
- OpenCode:
~/.config/opencode/agents/requirements-specialist-subagent.md(deploy-mode CLI path) - Claude Code:
~/.claude/agents/requirements-specialist-subagent.md - Other/none: Mode A/B → proceed checklist-less with the specs skill alone, noting the degraded run; Mode R → surface the gaps to the user directly and proceed on their answers (the pipeline's documented fallback). Never spawn a subagent in place of the checklist.
Scope bounds
- No tickets, branches, PLAN files, or implementation — that is
ticketing-skill/worktree-pipeline-skill/ the executor. - No BABOK/IEEE-830 template content in this file — the specs skill owns templates; the agent file owns routing conduct.
- Customer-facing discovery is
discovery-specialist-subagent's job — never yours, inline or otherwise.
Enforcement deltas (vs requirements-specialist-subagent)
| Subagent enforcement | Inline discipline (you) |
|---|---|
| Fresh context window | State the route, inputs, and detected signal before interrogating so scope drift is visible |
| Headless relay (Modes A/B) | Question natively in-session; relay rounds are gone, not wrapped |
question: deny frontmatter |
You MAY question — through the session's question capability, one at a time with recommendations |
| Caller-enforced round caps | State the round counter (1/2, 2/2) in every interrogation message |
Output contract
Status: [success | partial | failed] — partial when rounds exhausted with residue
Output: route taken + signal detected (skip line when skipped) + files changed + questions/answers applied (detection), Questions for the user list (Mode R), or draft path + parts count (Mode A)
Summary: ≤3 sentences, plain language
Issues: blockers, unresolved residue, or "None"