Retentioneering Product Analytics Agent Skill

Retentioneering product analytics

Objective

Turn event-level behavioral data into a reproducible answer to a product question —

why users convert, churn, loop, or abandon — using user trajectories, transitions,

funnels, and behavioral segments.

Do not merely generate visualizations. Connect each output to the question, separate

observation from interpretation, and never present path correlations as causal effects.

Bundled references (read on demand, not upfront)

| File | Read it when |

|---|---|

| references/api-map.md | before writing any Retentioneering call — verified signatures, argument conventions, return shapes for 5.x |

| references/analysis-recipes.md | after the question is clear — 10 field-tested patterns (R1–R10) with skeletons and pitfalls |

| references/gotchas-and-validation.md | before executing (API gotchas G1–G10) and before presenting (integrity checklist B1–B10) |

| scripts/inspect_event_log.py | step 2 — automated data profiling and schema suggestion |

Required event-log semantics

Minimum: a path identifier (user or session), an event name, a timestamp (or a reliable

order column — see gotcha G2 for order-only data). Useful extras: session id, segment

attributes (device, source, plan), event properties, conversion labels.

Workflow

1. Environment

• Confirm the package: python -c "import retentioneering; print(retentioneering.__version__)".

Expect 5.x; this skill's API map is version-verified for 5.0 — on a different major

version, trust installed docstrings over the map.

• Locate the event data (CSV / Parquet / frames in existing code). Never modify inputs.
• Do not invent methods: anything not in references/api-map.md must be verified

against the installed package before use.

2. Inspect the data BEFORE choosing methods

Run scripts/inspect_event_log.py <path> [--sep ...] (or replicate its checks inline for

in-memory frames). It profiles columns, infers the user/event/timestamp mapping, checks

timestamp parseability, duplicates, per-path ordering, path-length distribution, and

emits artifacts/data-profile.json plus a ready-to-paste Eventstream(...) schema.

Report to the user before proceeding: inferred mapping, row/user/event-type counts,

covered period, and any red flags (nulls in key columns, timestamp ties, suspected bots

or ultra-long paths, order-only timestamps). Confirm the mapping if inference is

ambiguous.

3. Frame the product question, then pick the SMALLEST recipe

Map the question to a recipe in references/analysis-recipes.md:

navigation structure/loops → transition graph (R1/R6) · before/after an anchor →

step matrix (R3) · ordered conversion flow → funnel (R1/R2) · what winners do

differently → diff on a funnel-stage segment (R2) · heterogeneous users → clustering

without target leakage (R5) · between two funnel levels → truncate micro-journey (R4) ·

intervention timing → time-to-outcome (R7) · cross-segment scan → segment overview (R8) ·

value of a fix → Markov what-if (R9, advanced).

Combine recipes only when each addition resolves a distinct uncertainty.

4. Execute reproducibly

• Prefer a rerunnable script (or a notebook executed top-to-bottom) over ad-hoc cells.
• Write artifacts to a dedicated output directory (artifacts/ by default).
• Log every filtering rule and its row/path impact; never silently drop data

(integrity item B2).

• Use sample_paths(frac=, random_state=) for stable subsamples; stochastic steps get

explicit seeds.

• Record lineage: save processed.recipe() and the package version into

artifacts/run-metadata.json — any artifact must be regenerable from raw data via

Eventstream.from_recipe(raw_df, recipe).

5. Validate before presenting

Work through references/gotchas-and-validation.md section B. Non-negotiables:

every percentage names its denominator; population filters are disclosed with counts;

survivorship and exposure confounds addressed; no outcome leakage into features;

small cells flagged with n; caption numbers come from headless *_data twins;

visuals agree with tables.

6. Interpret and deliver

Structure the final answer as:

• Observed — numbers with denominators and n.
• Interpretation — what it likely means.
• Alternative explanations — selection, structure, censoring.
• Product hypotheses — each with the metric an experiment would move.
• Suggested next analyses / A-B tests.
• Limitations.

Deliverables: analysis script or executed notebook; artifacts/data-profile.json;

artifacts/metrics.csv (key tables); interactive HTML exports via

widget.export_html(..., title=, analysis=) — write analysis= captions AFTER

conclusions are final; artifacts/summary.md (mapping, filters, assumptions, versions,

findings, limitations, next steps); artifacts/run-metadata.json (versions, parameters,

seeds, recipe() lineage).