원클릭으로
task-appraisal
How pre-mortems and post-mortems work, what verdicts mean, and how to act on appraisal results.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
How pre-mortems and post-mortems work, what verdicts mean, and how to act on appraisal results.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
How to orchestrate work across large documents and multi-agent delegations without burning token budgets. Reconnaissance-first, search-then-read, parallel-after-reconnaissance.
Decision criteria for delegating to the writer agent vs synthesising directly in your own response.
When and how to use the document library
How the Coder agent drives an OpenCode session via dispatch_coder — the engineering loop, when to dispatch, and how to verify the result is real (vs. claimed).
Maintain a personal index of interesting things, open questions, and follow-ups across sessions
When and how to render a promoted export to PDF via the export_pdf tool. Reads the writer's draft → export → PDF lifecycle and the operator-facing failure modes when the host is missing pandoc or tectonic.
| name | task-appraisal |
| description | How pre-mortems and post-mortems work, what verdicts mean, and how to act on appraisal results. |
| agents | cognitive |
The system automatically appraises your tasks at two points:
Pre-mortem — fires when you activate a task (if it's complex enough). Imagines the task has failed and identifies: predicted failure modes, blind spot assumptions, dependencies at risk, and information gaps. This is stored on the task and feeds the Forecaster's health evaluation.
Post-mortem — fires when a task completes, fails, or is abandoned. Evaluates whether the goal was achieved, compares pre-mortem predictions against reality, extracts lessons learned, and creates a CBR case from the outcome. A sensory event reports the verdict.
Neither blocks you. Both run as fire-and-forget processes after the lifecycle transition. If they fail (LLM error, timeout), the task proceeds normally.
| Transition | What fires | Condition |
|---|---|---|
| Task activated | Pre-mortem | Complexity >= medium, OR 3+ steps, OR part of an endeavour |
| Task completed | Post-mortem | Always. Simple tasks get deterministic "Achieved". Medium+ gets full LLM evaluation |
| Task failed | Post-mortem | Always full LLM, regardless of complexity |
| Task abandoned | Post-mortem | Always full LLM — abandonment has the most to teach |
| Endeavour completed | Endeavour post-mortem | Synthesises across all task post-mortems |
Simple 1-2 step standalone tasks skip the pre-mortem entirely and get a lightweight deterministic post-mortem. This is by design — the cost of an LLM call isn't justified for trivial work.
Post-mortems produce an AppraisalVerdict:
| Verdict | Meaning | What to do |
|---|---|---|
| Achieved | Goal fully met | Nothing — the system created a Strategy CBR case |
| PartiallyAchieved | Goal partly met, gaps remain | Consider a follow-up task for the gaps. Check lessons learned |
| NotAchieved | Goal not met | Review contributing factors. Check if the approach was wrong or the goal was unrealistic. A Pitfall CBR case was created |
| AbandonedWithLearnings | Stopped early, but something was learned | The learning is the value. Check lessons — they feed future planning |
When a post-mortem completes, you'll see a sensory event like:
<event name="post_mortem" title="Research pricing — partially_achieved">
Task task-abc123 post-mortem: partially_achieved. Key lesson: Two of three
competitor sites required authentication.
</event>
This is your cue to decide:
get_task_detail to see the full
post-mortem. Consider a different approach if retryingWhen you're about to start a complex task, the pre-mortem predictions are
stored on the task. The Forecaster uses them as a baseline — if a predicted
risk materialises, the health score reflects it. You can view predictions
with get_task_detail.
When planning similar future tasks, recall_cases will surface Pitfall
cases from past post-mortems. Read the lessons before committing to an
approach.
Controlled by [appraisal] in config.toml:
min_complexity — minimum task complexity for pre-mortem (default: "medium")min_steps — minimum steps for pre-mortem (default: 3)model — LLM model for appraisals (defaults to task_model)max_tokens — max tokens per appraisal call (default: 4096)