ML-196: add application-task pi prompt template

Introduces /application-task for working on Backlog.md tasks end-to-end:
- Pre-flight: read docs, load task plan, load relevant skills
- Execution: present plan, short loops, log progress, handle scope
- Done vs Complete: finalize without archiving, per Backlog.md workflow

.pi/prompts/application-task.md

@@ -0,0 +1,68 @@
+---
+description: Work on an existing backlog task end-to-end
+argument-hint: "<TASK-ID>"
+---
+
+We're going to work on task $1 end-to-end. Follow the Backlog.md Task Execution workflow
+step by step, ensuring every step is completed before moving to the next.
+
+## Pre-flight
+
+Before touching any code:
+
+1. **Read the project documentation.** Open and read `docs/architecture.md` and
+   `docs/project-conventions.md`.
+
+2. **Load the task plan.** Use `backlog_task_view` to load task $1 and read its
+   implementation plan, acceptance criteria, and notes. Do not proceed until you
+   understand the plan.
+
+3. **Load relevant skills based on the task's domain upfront.**. When undecided
+   if a skill is relevant or not, opt for loading it anyway.
+
+## Execution
+
+Follow the Backlog.md Task Execution workflow:
+
+1. **Mark the task In Progress** and assign yourself via `backlog_task_edit`.
+2. **Present the plan to the user** — summarize the approach and ask for
+   confirmation before writing any code. Wait for explicit approval.
+3. **Work in short loops:** implement a step, run the relevant tests, and
+   immediately check off acceptance criteria with `backlog_task_edit`
+   (`acceptanceCriteriaCheck` field) when they are met. If you catch yourself
+   running in circles, stop and let the user know.
+4. **Log progress** with `backlog_task_edit` (`notesAppend` field) to document
+   decisions, blockers, or learnings.
+5. **If the plan needs to change**, update it first via `backlog_task_edit`
+   (`planSet` or `planAppend`), get confirmation, then continue.
+6. **If new work appears outside the original acceptance criteria**, STOP and
+   ask the user before expanding scope. Never silently add AC or create
+   follow-up tasks.
+
+## Done vs Complete
+
+When implementation is finished, do NOT call `backlog_task_complete` or
+`backlog_task_archive`. Those move the task to the archive — they are for
+periodic batch cleanup, not for marking work finished.
+
+Instead, follow the Task Finalization workflow:
+
+1. **Verify all acceptance criteria** are checked (use `backlog_task_view` to
+   review).
+2. **Verify all Definition of Done items** are checked.
+3. **Confirm tests pass** — run the full test suite or the relevant test files
+   and verify no new warnings or regressions.
+4. **Write the Final Summary** via `backlog_task_edit` (`finalSummary` field).
+   Include what changed, why, tests run, and any risks or follow-ups.
+5. **Update the plan** if the executed approach deviated from what was recorded.
+6. **Set status to "Done"** via `backlog_task_edit`.
+
+Only then is the task truly done.
+
+## After Finalization
+
+- **Never autonomously create or start new tasks.**
+- If follow-up work is needed, present the idea to the user and ask.
+- If this was a subtask and the user instructed you to work on a parent "and all
+  subtasks," proceed to the next subtask. Otherwise, ask the user whether to
+  continue.

backlog/tasks/ml-196 - Add-pi-template-to-work-on-application-task.md

@@ -1,9 +1,10 @@
 ---
 id: ML-196
 title: Add pi template to work on application task
-status: To Do
+status: Done
 assignee: []
 created_date: "2026-05-22 09:40"
+updated_date: "2026-05-22 14:12"
 labels: []
 dependencies: []
 ordinal: 33000
@@ -17,3 +18,48 @@ ordinal: 33000
 - Solicit loading relevant skills
 - Make sure it'd done, and not complete
 <!-- SECTION:DESCRIPTION:END -->
+
+## Acceptance Criteria
+
+<!-- AC:BEGIN -->
+
+- [x] #1 Template file exists at .pi/prompts/application-task.md with correct frontmatter (description + argument-hint)
+- [x] #2 Template instructs agent to read docs/architecture.md and docs/project-conventions.md before writing code without summarizing their contents
+- [x] #3 Template instructs agent to load relevant skills based on their descriptions and triggers, without hardcoding a domain→skill mapping
+- [x] #4 Template distinguishes Done status from task_complete/task_archive per Backlog.md finalization workflow
+- [x] #5 Template covers the full Backlog.md execution workflow: pre-flight → present plan → work loops → finalization → after finalization
+<!-- AC:END -->
+
+## Implementation Plan
+
+<!-- SECTION:PLAN:BEGIN -->
+
+Create `.pi/prompts/application-task.md` that:
+
+1. **Solicits architecture reading** — instructs the agent to read `docs/architecture.md` and `docs/project-conventions.md` before touching code
+2. **Solicits loading relevant skills** — maps task domains (database, UI, testing, workers, APIs, etc.) to their corresponding `.agents/skills/*/SKILL.md` files and instructs the agent to load all relevant ones
+3. **Distinguishes done from complete** — follows the Backlog.md Task Finalization workflow: verify AC, verify DoD, run tests, write Final Summary, set status to "Done". Explicitly warns against calling `task_complete` or `task_archive` for marking work finished.
+4. **Follows Backlog.md execution workflow** — mark In Progress, present plan, work in short loops, log progress, handle scope changes, finalize properly
+
+The template uses `argument-hint: "<TASK-ID>"` for autocomplete compatibility with the `/application-task <ID>` invocation pattern.
+
+Updated pre-flight after review: point 1 no longer summarizes doc contents, point 3 no longer hardcodes a domain→skill mapping — instead instructs agent to load skills based on their own descriptions and triggers.
+
+<!-- SECTION:PLAN:END -->
+
+## Final Summary
+
+<!-- SECTION:FINAL_SUMMARY:BEGIN -->
+
+Created `.pi/prompts/application-task.md` — a pi prompt template for working on Backlog.md tasks end-to-end.
+
+**What it does:**
+
+- **Pre-flight phase**: instructs the agent to read `docs/architecture.md` and `docs/project-conventions.md` before writing code, load the task plan from Backlog.md, and identify + read all relevant `.agents/skills/` files based on the task's domain (database, UI, testing, workers, APIs, etc.)
+- **Execution phase**: follows the Backlog.md Task Execution workflow — mark In Progress, present plan for approval, work in short loops, log progress, handle scope changes
+- **Done vs Complete distinction**: explicitly warns against `task_complete`/`task_archive` and follows the Backlog.md Task Finalization workflow (verify AC, verify DoD, run tests, write Final Summary, set status to "Done")
+- **After finalization**: reminders about not autonomously creating tasks and proper subtask handoff
+
+**Template**: invocable as `/application-task <TASK-ID>` via pi's prompt template system.
+
+<!-- SECTION:FINAL_SUMMARY:END -->