Consultation
Turn a vague ask — "ClientCo needs a new landing page", "we should refresh ClientCo's competitor set", "build me a skill that does X" — into a sharp, reviewed spec and a clean handoff, by interrogating the ask one question at a time and mapping what's known vs unknown.
This is the elicitation front-end our intake surface was missing. /discovery preps a human's prospect call; /proposal produces a doc from discovery notes; this skill is the interactive loop between a vague ask and a sharp spec.
When to run
Invoke when:
- "Scope this" / "turn this into a spec" / "help me think through [X]"
- A client or internal ask is genuinely vague or high-stakes and would otherwise surface gaps halfway through execution.
- You're about to enter plan mode on something underspecified.
Do NOT invoke when:
- The ask is already sharp — a one-line clarification does the job. Skip the loop.
- You need discovery-call prep (question bank for a human to ask a prospect) →
client-discovery. - You need the proposal document itself →
client-proposals. - You're pressure-testing a business idea for viability →
business-brainstorm.
How it composes (do not reinvent these)
This skill is deliberately thin — it wires four existing Genesys systems into an elicitation loop:
| Stage | Delegates to | Why |
|---|---|---|
| Question shape | AskUserQuestion |
Already the "2–4 options, one Recommended, one 'other'" convention |
| Inspect-before-asking preflight | brain-first-lookup |
recall → grep client folder → path read, before pinging the user |
| Adversarial review of the spec | /premortem --plan (+ scope-guardian-reviewer for client-facing scopes) |
Our review discipline is already hook-enforced — do NOT ship a competing 5-round reviewer |
| Spec format | planning-doctrine |
The spec is a plan; use the canonical structure, not a bespoke one |
If you find yourself writing a bespoke reviewer or a new plan format, stop — that's the anti-pattern this skill exists to avoid.
Inputs
| Input | Required? | Source |
|---|---|---|
| The vague ask | required | User provides (client request or internal build idea) |
| Client / target | recommended | Names the folder + brand voice to inspect |
company-context output |
recommended | Skips re-research for client scoping |
If the ask is missing, ask for it in one line. If it's already fully specified, say so and skip to the handoff.
The loop
resolve working doc → questioning loop (gap-map) → draft spec → review (premortem) → handoff
- Resolve the working doc. Client ask →
projects/consulting/active/{client}/docs/MMYY-{topic}-consultation.md. Internal ask → a session scratchpad (or~/.claude/plans/if it's heading to plan mode). If a matching doc exists, ask continue-or-new. Seed it with the four-quadrant board below. - Run the questioning loop (details below) until no material unknown remains except client-accepted assumptions.
- Draft the spec in
planning-doctrineshape (Context → Goals → Outcome → Audience → Strategy → Use cases → Steps → Premortem → Resources). - Review — route the spec through
/premortem --plan(andscope-guardian-reviewerif it's a client-facing scope). On BLOCKERS, never revise silently: bring each gap back as a numberedAskUserQuestion, update the board, redraft, re-review. Cap at 3 rounds, then present what remains and let the user decide. - Print the handoff — the delimited block below, with the real absolute spec path.
The four-quadrant gap board (the working doc)
The consultation doc maintains these named sections, updated before each next question:
- Brief — the ask in one paragraph, as currently understood.
- Known knowns — stated facts and constraints.
- Unknown knowns — implicit taste and preferences, surfaced by showing options (this is the move — when taste is hard to verbalize, show 2–4 concrete references and let the choice reveal the preference; lean on the taste-library / a1-gallery / brand-kit for real artifacts).
- Known unknowns — named open questions.
- Unknown unknowns — blind spots you surface through expertise,
brain-first-lookup, or inspecting the target territory (the client folder, the live site, the existing skill). - Decisions — what's been settled, dated.
- Next step — a copy-paste next action, so any sitting can resume.
Pre-seed the gap board. When the ask targets a known deliverable type, seed Known-unknowns with that deliverable's required inputs from the ontology — e.g. scoping a landing page pre-seeds "positioning locked? messaging locked? brand-kit present? primary CTA? proof points?"; scoping a competitor refresh pre-seeds "which competitors? which dimensions? aggregate or per-competitor?". The gap board then partly drains as a checklist.
The questioning loop
Repeat until no material unknown remains except client-accepted assumptions:
- Pick the highest-leverage gap on the board.
- Inspect before asking — run
brain-first-lookup(recall → grep the client folder → read the path) to see if a fact or the territory already answers it. Don't ask the user what the workspace already knows. - Ask one question via
AskUserQuestion— a single numbered question with 2–4 options, one clearly marked (Recommended), plus the implicit "Other". Lead with options and your recommendation; the client owns taste and the final call. - Update the board — quadrants, Decisions, Next step — before the next question.
Rules: one question at a time (never batch a wall of questions); stop at the handoff, never start building; end every sitting with the doc current and a copy-paste next step.
The spec (draft output)
The spec follows planning-doctrine canonical structure — not a bespoke format. Map the consultation into: Context (SQCA) · Goals · Outcome · Audience · Strategy · Use cases (named engagement + present-tense pain, per the use-cases doctrine) · Steps · Premortem (≥2 failure modes) · Resources. Client-facing scope docs additionally obey doc-output-structure.md.
The handoff prompt
After review passes, print this verbatim (replace the placeholder with the real absolute path):
---HANDOFF PROMPT START---
You are the execution session for the approved spec at `<absolute path to spec>`.
Read the spec and every reference it names. Implement it while preserving its
outcome, decisions, constraints, out-of-scope items, and open assumptions.
Inspect the target territory (client folder, live site, existing skill) before
changing it via brain-first-lookup; ask only if the spec leaves a true blocker.
Follow the relevant Genesys voice + doc rules, run /premortem --output before
ship, then report what changed, what passed, and any remaining risks.
---HANDOFF PROMPT END---
For an internal build that heads straight to plan mode, the "handoff" is simply: enter plan mode with this spec as the plan file.
Self-roast (run before handoff)
- Anti-hallucination: mark unknowns
[PLACEHOLDER: …]; never invent a client constraint, metric, or preference the consultation didn't surface. - Composition check: review ran through
/premortem, not a bespoke reviewer; the spec is in planning-doctrine shape, not a home-grown format; questions usedAskUserQuestion. - Quality pass: the four quadrants are current; every Decision is dated; the Next step is copy-paste; the seven-tenet gate (
output-tenets.md) passed. - Skip check: if the ask was already sharp, did I avoid adding ceremony?
Integration with other skills
| Skill | Relationship |
|---|---|
| company-context | Inspect-before-asking source for client scoping |
| client-proposals | A locked client scope feeds the proposal |
| premortem | Owns the adversarial-review stage (composed, not reinvented) |
| business-brainstorm | Sibling for idea validation (this skill scopes an already-decided ask) |
Attribution
The consultation-elicitation loop — the four-quadrant Known/Unknown gap board, one-question-at-a-time questioning, "surface implicit taste by showing options," "inspect before asking," "lead with options + a recommendation," "never revise silently," and the delimited handoff block — is adapted from the consultant skill in eliasstravik/skills (MIT, © Elias Stravik, accessed 2026-07-13). Concepts re-voiced; the review + plan-format stages are Genesys-native. See .claude/discovery/0726-eliasstravik-skills-steal-analysis.md.