Identify three kinds of interruption before installing anything
A small theater is repairing its performance-catalog preview. A poster scan needs a person to confirm missing information, the agent repeatedly reads irrelevant files in a large directory, and an external summary request sometimes fails. Use three fictional performances and local response fixtures to practice notifications, context organization and debugging. Success means knowing why work stopped, resuming from traceable context and reproducing a repair.
The original topic groups user experiences into three useful scenarios. Its user counts, download totals and time-saving claims are not verification results for this workflow. The procedures below produce inspectable artifacts without presenting unverified popular names as confirmed SkillPI catalog entries.
Notifications: distinguish completion from waiting for input
Give each scan a task identifier and record four states: running, waiting for confirmation, succeeded and failed. When a poster has no year, produce a confirmation item and stop dependent work. Other authorized checks that do not depend on that year may continue. A process exiting successfully and a conversation waiting for a person are different events; a single generic completion message obscures the distinction.
For Claude Code, consult the current official hooks reference before choosing event names, input fields and execution locations. The official automation patterns on GitHub are a useful starting point for event routing. Their operating-system examples are not universal terminal scripts. Check Codex or another host independently instead of copying another product's settings path.
Start with a minimal local notification containing the task identifier, short project label, whether input is needed or execution ended, and a direction to return to the terminal. Do not put credentials, complete commands or unpublished theater schedules into popups. Notify once for a repeated state, keep the event time and last sent state, and clean up subscriptions or processes when the task ends. A notification must not approve a permission request or supply an answer on a person's behalf.
Exercise three cases: a confirmation request produces a waiting alert, a completed scan produces a success alert, and a simulated scan failure produces a failure alert linked to the correct local log. Also test disabled system notifications, duplicate events and a closed terminal. Keep a state table, event mapping and rehearsal record. If background delivery is unavailable, retain an explicit local status check and describe that limitation rather than promising that polling will never be needed.
Context: reduce rework in a large project
Do not load thousands of images, all historical logs and every guide into the initial conversation. Begin with the performance manifest, naming rules, scan entry point, three samples and acceptance conditions. Ask the agent to name additional files it needs and explain why, then provide context along dependencies. Keep source images read-only and write scan outputs to a separate artifact directory so a failed attempt cannot overwrite the originals.
Separate durable rules from current task state. Rules say that an unknown year stays empty, guessing is prohibited, and each record identifies its source file. State says which samples were scanned, what was found and what happens next. Attach a file or revision reference to conclusions. Mark a superseded conclusion explicitly; persistent memory should not preserve a repaired mistake as current truth.
A handoff card could read: task T17; baseline A; P02 lacks a confirmed year; output contains three records; local parsing checks passed; external summaries have not run; next action is for the owner to provide the year or accept an empty value. In a new conversation, first ask the agent to summarize that state and identify its evidence. Continue only after checking that it read the current version.
Measure time to finish the same accepted fixture, repeated reads, avoidable rework and actual charges. A shorter prompt can cost more if missing rules cause multiple attempts. Aim for the smallest sufficient context. Consider parallel work only when inputs and outputs are explicit, permissions cover the work and tasks do not write the same files. This is workflow design advice, not a claim that agents were created or run for this article.
Debugging: locate the failing layer before retrying
Disable real external calls for the first exercise. Prepare local responses for valid JSON, a missing field, a timeout and an error. Record the task identifier, stage, time, error category and a sanitized request identifier. Never store credentials or complete private prompts in the diagnostic bundle. Locate the failure in poster reading, structure parsing, the network request or page rendering. Those layers need different repairs.
For requests sent directly to the Claude API, use the official error documentation to distinguish authentication, authorization, request format, rate limits and service overload. A status code is a starting signal: inspect the error type, request identifier and environment too. If another provider or proxy produced the response, consult that system's documentation rather than applying the same explanation automatically.
Fix invalid credentials and malformed requests before retrying. Consider bounded backoff for transient failures, checking whether the client already retries so two mechanisms do not multiply attempts. Set an attempt limit, total time budget and stop condition. For operations that can write data, establish idempotency and the previous outcome before resending. This read-only practice scenario does not demonstrate that every production request is safe to retry.
The preview should display a temporary summary-unavailable message while preserving successfully scanned performance details. Add regression cases for recovery updating only the summary area, an unconfirmed year remaining empty and the retry button reflecting its real state. Deliver reproduction steps, a diagnosis by layer, a minimal patch, failure and recovery evidence, and explicit boundaries around the live service that was not tested.
Skill selection: Finder first, then verify the target
Check whether SkillPI Skill Finder is already installed. If missing, follow its official installation instructions. Describe the specific task before searching. Inspect the upstream version, full SKILL.md, license, dependencies and write scope of each candidate. A familiar name such as fast-claude-code does not establish that its SkillPI entry has been verified. Do not assume all skills are free, consume no resources or install identically across hosts.
The verified match here is Webapp Testing, used for the third scenario's local failure-and-recovery page checks. Its Python Playwright workflow requires Python, available browser dependencies, a local server startup command, a port and test fixtures. Observe the page structure before writing stable locators and assertions. Retain the script, necessary screenshots and browser logs with the revision tested.
This skill does not manage system notifications, guarantee memory across sessions or diagnose every API. It does not replace parser unit tests. When no notification-specific catalog skill has been verified for the task, use the host's documented notification mechanism instead of linking an unrelated card. As new skills become available, compare them against the same failure and recovery fixtures before updating the association and verification date.
Troubleshooting and FAQ
If a notification never arrives, first establish that the expected event occurred. Then inspect handler logs, execution environment and notification permissions; do not begin by disabling security controls. For duplicates, inspect deduplication identifiers and cleanup. An alert appearing on screen does not prove the underlying task succeeded.
If an old mistake returns after a new conversation, check that the current handoff was loaded, the rule has a concrete evidence reference and obsolete state was replaced. A vague reminder to avoid mistakes does not define acceptable input and output.
If an installed skill is missing, inspect the selected host's discovery location, metadata, naming conflicts and reload requirements. Validate with one minimal task. Do not invent a percentage suggesting that the path is always the cause.
Must every category be installed? Start with the interruption that matters most and expand only after its artifacts improve. Account separately for package price, model calls, external services and maintenance time. This exercise uses simulated data and makes no claim to have tested the reader's device or verified community-user totals.