update
May 10, 2026
By Teun
Open-source skill forces coding agents to prove completion
agent-postmortem-skill is an open-source verification skill that makes coding agents back up completion claims with evidence. It checks git state, command output, and exit codes before a task can be marked done.
agent-postmortem-skill is an open-source verification skill for coding agents that requires evidence before a task can be marked complete. The project is designed to catch cases where an AI assistant says it is finished, but the files were not changed, tests were never run, or a command failed and the agent kept going anyway.
The project was posted as a Show HN item by its creator, who described it as a way to turn “trust me” into “show me.” According to the project page, the skill works with any coding agent that can run shell commands and read git state.
The core idea is simple: do not accept a confident final summary without proof. The skill is meant to prevent “fake-done” states from reaching a branch by making the agent collect evidence before it can declare success.
According to the project description, the tool standardizes completion quality across both humans and agents. It also creates a portable postmortem artifact that can be reviewed, shared, and audited after the work is done.
Under the hood, agent-postmortem-skill enforces a completion pipeline with four steps. First, it captures an intent snapshot, which records the requested outcome and the success criteria. Then it collects evidence such as git status, git diff, command outputs, and exit codes.
Next, the skill compares the claim against the evidence. If the evidence is missing or if the commands show failure, the task is not considered complete.
The final step is a postmortem output that includes the verdict, proof, unresolved risks, and next actions. That report is meant to give users a clearer record of what actually happened, rather than relying on the agent’s own summary.
The project’s quickstart instructions say to clone the repository from GitHub and copy SKILL.md into the agent’s skill directory. The source did not include the full directory examples, but it did say the skill is open source and ready to be dropped into supported agent setups.
For teams using AI coding agents, the project is aimed at reducing hidden failures that only show up later in review or deployment. It does that by making completion dependent on verifiable signals, not just the agent’s narrative of success.
