Connect skills with handoff contracts
A bakery wants clearer collection instructions for preorders. Customers mistake preparing for ready to collect, while staff worry that premature reminders create more explanations. The PM wants to connect feedback, data and requirements instead of reorganizing the same material every week. Use synthetic comments, interview excerpts and practice order events to define three handoffs: a research findings table, a metrics-and-decision record, and a PRD/reporting package.
Each artifact states its version, owner, input scope, evidence locations, unresolved questions and acceptance conditions. Downstream work reads only an agreed version. An unsupported conclusion remains a question even after passing through several agents. This teaching workflow does not claim to access a shop system, contact customers or send notifications.
Handoff one: research with traceable comments and interviews
Collect explicitly authorized feedback exports and describe channels, time range, selection rules and missing coverage. Public visibility does not imply unlimited scraping or permission for external transmission. Use supplied local samples when appropriate access is unavailable. Remove names, order identifiers, phone numbers and other identifying details, replacing them with research identifiers. Deduplicate repeated exports of one record without merging different people merely because their comments resemble each other.
Frame concrete questions: when do customers decide collection is possible, which cue determines arrival time, and what happens after an unexpected wait? Classify comments provisionally, requiring supporting identifiers and counterexamples for each theme. Audit classifications before counting sample frequencies. Did not see a reminder may mean delivery failure, a reminder not yet sent, a different channel or wording that omitted the next action.
For interviews, separate statements, behavior, need level, context, decision journey, motivational hypothesis and follow-up questions. These dimensions guide inquiry; they are not boxes that must be filled with invented psychology. Commuters and people collecting for relatives may need different information, but segmentation must follow evidence. Leave unknown motives unknown.
Accept the first handoff when needs trace to excerpts, counterexamples remain present, sample limits are stated and questions have owners. Finding F3 might describe confusion between preparing and ready to collect. Its candidate requirement is a state explanation, not an immediate commitment to SMS. Missing channel coverage limits the conclusion to the supplied material.
Handoff two: inspect data before specifying metrics
Inventory the available order-created, preparing, ready and collected events with timestamps, versions and provenance. Confirm timezone and field definitions, distinguishing orders, events and visits. Define treatment of repeated exports, missing collection time, test orders, cancellations and collection on a later day. Do not let a tool silently drop awkward records to produce a cleaner chart.
Use Exploratory Data Analysis at this step for a bounded field, missingness and aggregate-anomaly profile of an authorized local CSV. Core CLIs need Python 3.11 or later; the complete optional environment has additional dependencies. Read the pinned SKILL.md for exact scope. Bound input and output locations, preserve raw data and avoid identifying rows. Stop on unsupported formats. The skill is not a dashboard connector and does not establish experimental causation.
Define metrics beyond their names. Activation could provisionally mean the proportion of eligible practice orders completing a confirmation after the first view of collection instructions. If retention matters, define the observation unit and interval first; daily return is a poor automatic assumption for infrequent bakery customers. Specify numerator, denominator, exclusions, latency, filters and ownership. Hand-check ten synthetic records before expanding the analysis.
Write the experimental hypothesis before results: explaining that preparing is not yet collectible may reduce early arrivals. Define allocation units, primary metric, observation window, guardrails and stopping conditions. Check for customers crossing groups, holiday preparation differences and missing events. With insufficient observations, report descriptions and gaps rather than declaring a small numerical difference significant. Statistical conclusions require a suitable method and responsible review.
Turn a decision record into priorities
Compare state wording, estimated collection time and external reminders using evidence strength, reach, cost, risk and dependencies. In this scenario, start with wording and state explanation because understanding can be tested without adding a communication channel. Reliable estimated-time data is not ready, and external reminders need separate prerequisites.
The record includes the selected option, reasons, rejected alternatives, remaining hypotheses and review conditions. The agent organizes the material rather than deciding on the basis of installation counts. After owner agreement, create tasks to standardize state terms, draft normal/delayed/failure messages, inspect fixtures and decide whether to pilot. Each task has an owner, dependency, completion evidence and acceptance rule. As soon as possible is not a schedule.
Handoff three: PRDs and reports share an evidence chain
The PRD consumes agreed finding F3 and the decision version. Describe target users, the problem, scope, state flow, permissions, exceptions, acceptance and open issues. A sample rule says that preparing must not invite immediate arrival, ready may show collection instructions, and failure to retrieve status must display unknown with a support route rather than default to ready. These are practice states; engineering and shop owners must confirm the real system mapping.
Define fixtures with input state, expected wording, permitted action and prohibited action. Separate copy changes from real delivery logic; this iteration sends no messages. When a requirement changes, update fixtures and references together so reviewers can trace the rule to its research rationale and the PRD does not diverge from reporting.
Generate weekly reporting from task and acceptance records. Completed items cite evidence, unfinished items identify blockers and owners, and next-week plans remain future actions. Draft generation is not a release, and proposed metric definitions are not an operational dashboard. Adapt emphasis for shop staff, engineers and owners without giving different audiences contradictory facts.
At the technical-prose review step, Writing Guidelines can return located revision suggestions. Provide documents, audience and access to its linked handbook, then adapt recommendations to team language. It does not verify facts, sign into systems, retrieve data or export Word. For a formatted Word deliverable, use a suitable exporter and reopen tables, links and pagination before calling the format accepted.
Installation and report-pipeline prerequisites
Check whether SkillPI Skill Finder is installed before a target skill; if missing, follow its official instructions. Search by artifact and inspect source, version, license and execution requirements. The open-source PM Skills method library can inform staged artifacts without being represented as a verified SkillPI card. Consult official Codex skills documentation for invocation behavior. Similar descriptions can overlap, so coexistence is not a guarantee of conflict-free execution.
An automated report also requires system access, field definitions, conversion and delivery scope. Verify them separately. Use a business system only with appropriate authorization; this exercise starts from local fixtures and retains manual export if no suitable connector has been verified. Installing instructions does not create an account or grant data permissions.
Evaluate the first week before expanding
Start with one repeated stage and compare total time, corrections, omissions, useful findings and actual charges for the same materials. Keep justified rule changes and rerun old fixtures. When a new skill enters SkillPI, test it against this handoff contract before updating associations.
Does this work outside software companies? Provenance, metric definitions, decision records and acceptance transfer, but domain owners must supply business rules rather than inherit generic internet retention metrics. Can multiple skills conflict? Descriptions, dependencies and formats may overlap; specify each step's inputs, outputs and owner. What is needed first? An authorized sample and a clear acceptance task reveal the real bottleneck more reliably than installing six names at once.