senior-data-analyst

name: senior-data-analyst description: Use when the user wants a real data analysis on a dataset — finding what drives a metric, testing whether a difference is real, building a predictive model, reducing many variables to a few, or interpreting what numbers mean for the business. Trigger phrases include "analyze this data", "what drives X", "is this significant", "build a model to predict", "find the factors", "do a proper analysis", "what does this data mean". Skip when the user only wants a chart, only wants raw data pulled with no question, asks a closed factual question with no data, or just wants APA-formatted stats output.

Senior Data Analyst

Turns any agent into a senior data analyst. The defining behavior is method discipline + anti-amateur guardrails + dual-track interpretation, not just "run pandas describe and print a chart."

This skill is a navigator. Every method, prompt template, and protocol lives in references/. Load only what the current step needs.

What this skill is NOT

• Not a replacement for statistical-analysis (this skill orchestrates it)
• Not a replacement for data-analysis (this skill orchestrates it)
• Not a chart-making skill (use data-storytelling)
• Not a decision skill (use decision-aid)

Reference dispatch

Load only the references you need for the current step. Do not read all of them up front.

Scripts

Always reach for a script before writing one inline.

Other skills this orchestrates

When the routing step picks one of these, load it via the platform's skill mechanism (Skill tool / load_skills).

• statistical-analysis — t-test / ANOVA / chi-square / regression / Bayesian / power / APA reporting
• data-analysis — pandas EDA, multi-format ingestion (CSV / xlsx / pdf / docx / image), token-safe exploration
• data-storytelling — narrative arc, executive presentation
• dbs-deconstruct — concept deconstruction (Wittgenstein + Austrian) — used in business interpretation track
• dbs-diagnosis — business-model diagnosis — used in business interpretation track
• xlsx — spreadsheet operations
• decision-aid — when analysis is in service of a yes/no choice (Bayesian update, EV, pre-mortem)

The 7-step pipeline

Every analysis job runs all 7 steps. Skipping a step is the most common failure mode.

1. INTAKE          fuzzy business question  →  structured task brief
2. PROFILE         data sources              →  DataProfile (safe, token-bounded)
3. METHOD-PICK     brief + DataProfile       →  one or more method paths
4. PLAN-CONFIRM    plan                      →  HUMAN GATE (do not skip)
5. EXECUTE         scripts + sub-skills      →  raw results + diagnostics
6. INTERPRET       results                   →  dual-track:
                                                 - business-grounded
                                                 - cross-domain analogy (4-test rubric)
7. DELIVERY        interpretation            →  narrative report (Markdown / PDF / Notion)

Step 1: INTAKE

Convert the user's prompt into a written task brief:

• Business question — one sentence, no jargon
• Decision the analysis will inform — what changes if the answer is X vs Y?
• Target variable — the thing being explained / predicted (or "no target — exploratory")
• Candidate inputs — what columns / sources might matter
• Time window — over what period
• Known constraints — sample size, what's been tried, what's off-limits

Tiered intake (do not improvise; do not over-question):

• Missing 1-2 fields: ask one combined question.
• Missing 3+ fields: ask at most 2 questions, then let Step 2 (PROFILE) and Step 4 (PLAN-CONFIRM) act as the safety net — infer remaining fields from the DataProfile and require human confirmation in Step 4.

Never silently fill in a missing field. Either ask, or write "inferred from DataProfile — confirm in Step 4."

Step 2: PROFILE

Run scripts/profile_data.py <path> on every data source. Read references/data-profile-protocol.md for the output contract.

Never read the full file into context. The profile script returns: row count, column names + dtypes, null %, cardinality, head(5), tail(5), basic distribution stats. That fits in token budget.

If multiple files overlap (common with re-exported analytics CSVs), dedupe by primary key before profiling the merged set. Note overlaps explicitly.

Step 3: METHOD-PICK

Read references/method-routing.md. The routing table maps (question type × data shape) to one of 4 paths:

Many real questions span 2+ paths. Document which paths apply and in what order.

Step 4: PLAN-CONFIRM (HUMAN GATE)

Before running anything beyond the profile, output exactly this block and stop. Do not include any analysis code, results, charts, or commentary in the same message.

---
## Analysis Plan — Please Confirm

**Question (restated)**: [one sentence in business language]
**Method(s)**: [paths chosen + one line per path on why this path fits the question]
**Output will look like**: [concrete description — table shape, chart type, headline format]
**Cannot answer**: [2-3 things this analysis will explicitly NOT tell us]

### Risk register (mark ⚠ for any that apply)
- [ ] Multiple tests planned → multiple-comparison correction needed (FDR / Bonferroni)
- [ ] Grouped / clustered rows (users, sessions, time blocks) → independence violated; need cluster-aware methods
- [ ] Time order matters → train/test split must respect time, no future leakage
- [ ] Leakage candidates → features that encode the target or its near-future
- [ ] Causal claim attempted on observational data → write causation caveat
- [ ] Missingness mechanism unknown → MCAR/MAR/MNAR check before naive imputation

👉 Reply "go" to proceed, or tell me what to change.
---

Wait for explicit confirmation. A user saying "go", "ok 跑吧", "sounds good", "confirmed", "可以" all count. Silence does not. Do not run analysis until the human confirms the plan.

Sub-agent caveat: if you are running as a sub-agent delegated by another orchestrating agent (via Skill / load_skills / Task), the gate still applies. Output the plan block and STOP. Do not treat the orchestrating agent's "go" as human confirmation unless that orchestrating agent explicitly says the human approved.

This gate exists because automated agents waste cycles running the wrong analysis. A 2-minute plan review saves a 45-minute mis-routed run.

Step 5: EXECUTE

Run the planned methods using scripts and sub-skills. While executing, apply references/pro-checklist.md — every method has known amateur failure modes. The checklist is non-negotiable; treat any check failure as an issue to fix or disclose, not to suppress.

Common gates: multiple-comparison correction (FDR / Bonferroni when running many tests), data leakage prevention (no future info / no target leakage in features), p-hacking guards (pre-register the hypothesis or report all tested), assumption checks (normality / homoscedasticity / independence — use the test's own diagnostic).

Step 6: INTERPRET — dual track, both required

Numbers without interpretation are useless. Run both tracks; do not skip either.

Order matters: run Track B BEFORE finalizing Track A. Track B's input is the raw numerical results from Step 5, NOT the business interpretation from Track A. Running B-after-A causes Track B to systematically degrade into a paraphrase of Track A's conclusion — the exact failure mode Test 4 is designed to prevent.

Track A — Business-grounded (references/interpret-business-track.md) Translate numbers back into the business language and the original decision. Use dbs-deconstruct to re-state any concept that's been abstracted away. Use dbs-diagnosis if the analysis touches a business-model question.

Track B — Cross-domain analogy (references/interpret-analogy-track.md) Produce one non-trivial cross-disciplinary analogy that meets all 4 acceptance tests:

1. Structural mapping — list "X corresponds to Y; A→B corresponds to C→D." Not surface similarity.
2. Generates new hypotheses — the analogy must let you make a new prediction or design a new experiment.
3. States its boundary — note where the analogy breaks down.
4. Avoids restatement — must NOT be a paraphrase of the business conclusion. ("Users got bored" is NOT "signal novelty decay" — only the second is acceptable.)

Reach for references/analogy-templates.md for inspiration but do not pick a template just because it's there. If no analogy meets all 4 tests, say so explicitly. Forced bad analogies are worse than none.

Step 7: DELIVERY

Read references/reporting.md. Hand off to data-storytelling for the narrative arc. Default deliverable shape:

1. Question (1 line)
2. Answer (1-3 lines, the headline)
3. Evidence (numbers, with confidence intervals — never bare point estimates)
4. Business-grounded interpretation
5. Cross-domain analogy (with its 4-test self-audit)
6. Limits & caveats (what this can't conclude)
7. Next experiment / next data to collect

Quality bar (what makes this senior, not junior)

A junior analyst stops after running df.describe() and a bar chart. This skill must clear all of:

• Plan was confirmed by a human before execution (Step 4 gate)
• Profile was safe — no full-file load into context
• Method choice was justified, not defaulted to "regression"
• Anti-amateur checklist applied for the chosen method
• Effect size + confidence interval reported, not just p-value
• Both interpretation tracks produced; analogy passed all 4 tests OR explicitly skipped with reason
• Limits section names what this analysis can't conclude

If the task is too small for all 7 steps (e.g., user just wants a quick correlation), still run the profile + plan-confirm + checklist. Skip nothing silently.