dead-letter

Dead Letter Skill

You are processing a failed task — one that returned an error, a blocker, a partial result, or a "couldn't proceed" message. Your job is to make sure nothing is lost, the failure mode is named, and the next action is unambiguous.

Failed tasks are information. The agent that failed learned something — what it tried, what broke, what it couldn't resolve. If that context isn't captured before the session ends, the next agent starts from zero and hits the same wall.

What You Will Do

1. Read the failed task output — the full message, error, or partial result from the agent that couldn't complete.

2. Classify the failure mode:

1. Write the dead letter record to .claude/dead-letter-<timestamp>.md:

# Dead Letter — <ISO timestamp>

## Task
[Original task, one sentence]

## Failure Mode
[CODE]: [one-line description]

## What was attempted
- [Step 1 tried]
- [Step 2 tried]
- [Where it broke and why]

## What was learned
- [Fact that wasn't in the original brief]
- [Constraint discovered during execution]
- [Any partial output or intermediate state that exists]

## Retry packet (if retryable)
**Fix required before retry:** [specific thing to change]
**Retry prompt:**
> [Revised task with root cause addressed]

## Escalation (if not retryable)
**Route to:** [Agent or human who can unblock]
**Decision needed:** [Specific question that must be answered]
**Deadline:** [When this blocks downstream work]

## Files written (partial output to preserve)
- `path/to/file` — [what's in it, what's missing]