Imported from borghei/claude-skills (research-ops/market-research/SKILL.md). Install upstream with npx skills add borghei/claude-skills --skill market-research. Copyright stays with the author (MIT + Commons Clause).

Market Research

Applied market research for people who have to defend a number in a room. This skill is about the operational craft: constructing a market size two independent ways, reconciling the gap, cutting the market into segments that behave differently, and fielding survey instruments that do not manufacture the answer you hoped for.

When to use this skill

• Sizing a market for a board deck, investor memo, or funding request where the number will be challenged line by line
• Reconciling a TAM you inherited — an analyst report says $12B, your bottom-up build says $700M, and you need to explain the gap
• Segmenting a market before a pricing, packaging, or GTM decision
• Triangulating demand signals (search volume, inbound, win rates, analyst data, competitor headcount) into one directional read
• Designing a survey to answer a market question — willingness to pay, category awareness, switching intent — without leading the respondent
• Auditing someone else's sizing before you sign off on it

Inputs the skill expects

• The market definition in one sentence — including geography and buyer
• A top-down anchor (published market value) with its source and vintage
• Bottom-up unit economics — unit count, qualified share, annual value per unit
• The decision the number is feeding (investment size, hiring plan, pricing)
• Time horizon for SOM (1 year vs 3 years changes it by an order of magnitude)
• For surveys: population size, target margin of error, mode (panel, list, intercept)

Clarify First

Before generating, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Market definition — what is in and what is out — the single biggest driver of the number; "dental software" and "dental practice management software for multi-chair EU practices" differ by 20x
• The decision this sizing supports — a fundraise tolerates a wide TAM; a hiring plan needs a defensible SOM
• Time horizon for SOM — 12-month obtainable share and 3-year obtainable share are different artifacts
• Whether a published anchor exists and its vintage — a 2022 report in a 2026 memo needs an explicit growth bridge

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Build and reconcile TAM/SAM/SOM

1. Write the market definition sentence first. Everything downstream inherits it.
2. Build the top-down chain: published market value, then named filters that each cut it (geography, segment, buyer qualification), each with a retention fraction and a stated justification.
3. Build the bottom-up chain independently: unit count from a countable source, qualified share, annual value per unit, reachable share, expected win rate.
4. Run the builder. It computes both chains, reconciles them layer by layer, and flags implausible ratios and divergence.
5. Resolve every fail before the number leaves your machine. A warn needs a sentence in the memo, not a fix.

python3 research-ops/market-research/scripts/tam_sam_som_builder.py \
  --input research-ops/market-research/assets/sample_market_model.json \
  --format text

Workflow 2 — Triangulate demand signals

1. Collect every observable demand signal you have — search volume, inbound lead velocity, win rate by segment, analyst growth rates, competitor hiring, category conference attendance.
2. Score each for source independence and directional strength.
3. Run the triangulator to get a weighted demand index and, more importantly, the list of signals that contradict each other.
4. Investigate contradictions before averaging them away. A conflicting signal is usually a segmentation boundary you have not drawn yet.

python3 research-ops/market-research/scripts/demand_signal_triangulator.py \
  --input research-ops/market-research/assets/sample_demand_signals.json \
  --format text

Workflow 3 — Audit a survey instrument before fielding

1. Draft the instrument with the market question stated at the top.
2. Run the auditor. It checks each item for leading language, double-barrelled phrasing, absolutes, unbalanced or over-long scales, and missing escape options.
3. Check the sample-size verdict — it computes required n from population, target margin of error, and confidence level.
4. Fix every fail, then re-run. Field only on a clean run.

python3 research-ops/market-research/scripts/survey_instrument_auditor.py \
  --input research-ops/market-research/assets/sample_survey.json \
  --format text

Decision frameworks

Which sizing method for which situation

Plausibility thresholds

These are the ratios the builder enforces. They are heuristics, not laws — but crossing one without an explanation in the memo is how sizing loses credibility.

Survey sample size at 95% confidence

Required n for a proportion estimate, finite population corrected. Use these as a sanity check on the auditor's output.

The jump from ±10% to ±5% quadruples cost for a band most market decisions do not need. [RECOMMENDED] Field at ±10% for directional category questions and reserve ±5% for pricing and packaging decisions where the band drives the choice.

Anti-Patterns

The Inherited TAM

Mistake: Copying a market size from an analyst report or a competitor's deck into your own memo, adjusting the geography, and presenting it as your build. Why it happens: The number is already large and already sourced, and building bottom-up takes two days you do not think you have. Instead: Use the published figure as the top-down anchor only, and always build the bottom-up chain alongside it. The reconciliation gap is the most informative artifact of the whole exercise — it tells you exactly which definition the report used and yours does not.

The Multiplication Fantasy

Mistake: SOM computed as "if we capture 1% of the TAM" with no mechanism behind the 1%. Why it happens: It sounds modest, so nobody challenges it, and it produces a convenient number without requiring a channel model. Instead: Build SOM from reachable units times expected win rate, where both come from something observed — your funnel, a pilot, or a comparable. If you cannot name the channel that reaches those units, you do not have a SOM.

The Stale Anchor

Mistake: A four-year-old market report used at face value in a current memo. Why it happens: It was the best available source when someone first built the model, and nobody re-checks a number that has been in the deck for a year. Instead: Record the vintage of every anchor. If it is more than 18 months old, apply an explicit growth bridge with a stated CAGR and show both the raw and bridged figures. An unbridged stale anchor invites the reviewer to discount everything downstream of it.

The Leading Instrument

Mistake: Asking "How valuable would an automated reporting feature be to your team?" and reporting the enthusiasm as demand evidence. Why it happens: The team already believes in the feature, and the question is written by the person who wants it built. Instead: Ask about the current behaviour and its cost — "How many hours last month did your team spend building reports manually?" — and let the demand fall out of the numbers. Run every instrument through the auditor before fielding; leading items are cheap to fix pre-field and impossible to fix post-field.

Segments That Do Not Behave Differently

Mistake: Cutting the market by company size or geography because that data is available, then finding every segment has the same conversion and the same ACV. Why it happens: Firmographic fields are in the CRM; behavioural ones are not. Instead: Segment on the variable that changes the buying decision — trigger event, existing tooling, regulatory obligation, or team structure. A segmentation is only useful if the segments have measurably different win rates or values.

Files