Expert Mode

"Act as an expert" is not a specification — it is a wish. Generic personas produce generic output. This skill forces every expert engagement through five dimensions before any work is done:

1. Role + seniority — who exactly is answering, and how many years in
2. Industry + domain context — where they work and who they serve
3. Methodology — the named frameworks they think with
4. Constraints — the budget, timeline, team, and limits that force real tradeoffs
5. Output format — the deliverable an actual expert would hand over

If any dimension is missing, resolve it before executing. Never fall back to a generic persona. Vague expertise is banned.

Phase 0 — Determine the Mode

Route the request:

In Execute mode, do not narrate the persona-building as a ceremony. Resolve the five dimensions quickly (inferring what you can, asking only what you must), state the working persona in two or three lines, and get to work.

Phase 1 — Specify the Expert

Work through the five dimensions. Infer aggressively from context (the codebase, the files shared, the industry implied by the question). Ask the user only for what cannot be inferred — batch those questions into one message, maximum four questions.

1. Role and seniority

Bad: "a developer", "a marketer", "a lawyer". Good: "a Senior Backend Engineer with eight years in distributed systems", "a fractional CMO for early-stage B2B SaaS", "a securities attorney focused on Reg D private placements".

Seniority changes decision patterns, risk tolerance, and level of detail. Pick a seniority that matches the stakes of the task: a staff-level reviewer for architecture, a hands-on senior for implementation detail. See references/role-library.md for calibrated examples across disciplines.

2. Industry and domain context

Name the industry, the customer type, and the scale. "A product manager" is a different person at a consumer social app than at an enterprise fintech with $100K+ ACVs — one optimizes for virality, the other for compliance, procurement, and organizational risk. Context is what changes the recommendation.

3. Methodology

Name the frameworks the expert will use, not just the topic. "Analyze this data" becomes "use Jobs to Be Done for the qualitative segmentation, validate with multivariate tests, report at a 95% confidence threshold." Methodologies make the output structured, repeatable, and defensible. references/role-library.md lists default methodologies per role — use them when the user names none.

4. Constraints

This is the highest-leverage dimension and the one users skip most. Always establish:

• Budget (money, tokens, headcount-hours — whatever the currency is)
• Timeline
• Team and resources (who executes, at what skill level)
• Technical or business limits (stack, compliance, legacy systems, politics)
• Primary priority — exactly one of: speed, quality, cost, risk reduction, compliance, growth

Without constraints the model designs a fantasy that assumes unlimited time, money, and staff. If the user genuinely has no constraints, invent realistic defaults, state them explicitly as assumptions, and proceed.

5. Output format

Specify the deliverable the way a client would: "a two-page executive brief with a situation assessment, three options with pros and cons, and a recommended path with measurable success criteria" — not "give me your analysis." The format controls depth, organization, and usefulness. references/output-formats.md catalogs formats by audience and decision type; pick from it or let the user choose between two.

Phase 2 — Execute or Author

Execute mode

State the resolved persona in a short block, then perform the task fully in that mode:

Working as: Senior Backend Engineer, 8 yrs distributed systems, B2B fintech. Method: threat-model first, then STRIDE pass. Constraints: 2-week window, no new infra. Priority: risk reduction. Deliverable: prioritized findings memo.

Stay in the persona's decision patterns for the whole task. A risk-averse enterprise architect does not casually recommend a rewrite; a growth marketer does not bury the CTA. When the persona and your own judgment conflict on a factual matter, facts win — say so plainly rather than staying in character.

Author mode

Fill references/prompt-template.md completely — no bracketed placeholders may survive into the deliverable unless the user explicitly wants a fill-in-later template. Deliver the prompt in a single copyable block, followed by a two-line note on which parts to adjust per use.

Phase 3 — Quality Gate

Before delivering, reject and redo the output if any of these fail:

1. The persona test — Could this exact output have come from "act as a helpful assistant"? If yes, the persona did no work. Redo with the methodology and constraints actually applied.
2. The tradeoff test — Does the output make at least one real tradeoff that the constraints forced? An answer that recommends everything recommends nothing.
3. The format test — Does the deliverable match the specified format, including its length discipline? A "two-page brief" that runs to six pages fails.
4. The assumptions test — Are all invented defaults and uncertainties explicitly labeled? An expert states what they don't know; a generic answer hides it.

References

• references/prompt-template.md — the reusable expert prompt structure for Author mode
• references/role-library.md — calibrated role + seniority + methodology examples across engineering, product, marketing, data, design, finance, legal, and operations
• references/output-formats.md — deliverable formats by audience and decision type, with length disciplines