Add second-opinion skill - #9
Open
gamellis wants to merge 5 commits into
Open
Conversation
Sends the whole session transcript, verbatim, to a different agent CLI and reports its answer back unedited. An authored briefing frames the question: it fixes which options exist and which premises are settled, so the second opinion can only choose between decisions already made. Passing the transcript hands over the frame itself, and in testing the other model challenged premises unprompted. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Flag syntax is CLI trivia the agent can recall or look up, and a table of flags rots. Keep only the three constraints it wouldn't default to: read-only, transcript in as the prompt, answer captured to a file. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Skill bodies arrive as user messages, so the extraction captured this
skill's own text. Its closing line — ask the user which harness — became
the transcript's last live instruction, and codex answered that instead
of the question. Observed on a real run.
Also close the transcript with the question in full rather than a bare
Answer Q{N}, which wasn't enough to outrank a trailing block of
instructions.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ablation on a real question, same transcript and model, one variable: with the clause the other harness listed three premises it disputed; without it the answer stayed inside the frame and audited nothing. Disagreement with the recommendation happened either way — challenging the question did not. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Puts the question you're stuck on to a different agent CLI — Codex, Cursor, or Claude Code — and reports its answer back verbatim, alongside where it diverges.
Why the transcript, not a briefing
The first version wrote a neutral briefing: the question, the background, the options, with the current agent's own recommendation deliberately withheld. That's worse than it sounds. A briefing you author is a briefing you frame — it fixes which options exist and which premises are settled, so the second opinion can only pick between decisions you already made.
Sending the raw transcript hands over the frame itself. Tested both ways on the same design question: the briefing run answered it well but accepted every premise, while the transcript run challenged three of them unprompted — including one that turned out to be load-bearing.
The fear that a transcript would anchor the other model, since it contains the current agent's recommendation and reasoning, didn't hold up. It diverged more, not less. A transcript reads as an argument to evaluate; a neutral briefing reads as a question to comply with.
What's in it
SKILL.md— export the transcript, pick a harness, relay the answer verbatim.transcripts.md— extraction recipes. The Claude Codejqfilter is written and tested; the Codex and Cursor entries point at the right stores and say to inspect before filtering, rather than guessing a shape.Deliberately minimal on instruction. Verification, frame challenge, and argument all happened without being told to — lines telling the agent to do them were no-ops. The one appended instruction that earns its place is which question is live, since that's the only thing the transcript can't supply.
Needs
jqand at least one other agent CLI. It ships the session to another vendor, so there's a privacy note in the skill.🤖 Generated with Claude Code