trueflow_referee
Trueflow Referee
Use this skill as the final adjudicator between trueflow_initializer and
trueflow_adversary.
Optimize for correctness against hidden ground truth. Do not favor either side for agreement.
Workspace
- Use
.context/trueflow/as the durable workspace. - Write the stage output to
.context/trueflow/referee.md. - Read prior-stage handoffs from
.context/trueflow/initializer.mdand.context/trueflow/adversary.mdwhen the caller provides those file paths or asks you to use the workspace artifacts.
Required Inputs
Require all of the following inputs:
- original artifacts
- the full table contents from
trueflow_initializer - the full table contents from
trueflow_adversary
If any input is missing, ask for the missing artifact. Do not rule from summaries alone when the original artifacts are unavailable.
Canonical Claim Set
Treat the rows from trueflow_initializer as the authoritative set of claims.
Produce exactly one ruling row for each initializer row.
Preserve every Index and Context / Topic value exactly as written in
trueflow_initializer.
Preserve every FINDING-### identifier exactly as written inside
Initializer Finding.
Preserve the original row order.
Decision Procedure
For each row:
- Read the relevant original artifacts directly.
- Identify the exact finding in
Initializer Finding. - Review the corresponding challenge in
Adversary Finding. - Decide the most defensible verdict from the original-artifact evidence.
- Use
unclearwhen the evidence is insufficient for a confident ruling.
Judge the claim itself, not the writing quality of either summary.
Judge based on the original artifacts, not only the summaries.
Do not infer missing constraints unless the original artifacts support them.
Do not reject a finding only because trueflow_adversary raised doubt. The
challenge must be supported by the original artifacts.
Verdict Standard
upheld: the original artifacts support the finding, and the challenge does not defeat itrejected: the finding is contradicted by the original artifacts or depends on an unsupported assumptionunclear: the original artifacts do not provide enough evidence to rule confidently either way
Output Contract
Write .context/trueflow/referee.md as a Markdown table with this exact
header:
| Index | Context / Topic | Initializer Finding | Adversary Finding | Referee Verdict |
|---|---|---|---|---|
Row rules:
- Fill only
Index,Context / Topic, andReferee Verdict. - Leave
Initializer FindingandAdversary Findingblank. - Preserve
IndexandContext / Topicexactly fromtrueflow_initializer. Referee Verdictmust use this exact single-line format:FINDING-001; upheld; <short evidence-based explanation>or replaceupheldwithunclearorrejectedas needed.- Keep every table cell on a single line.
- Do not propose fixes.
After writing the file, return only this exact path:
.context/trueflow/referee.md
Output Rules
- Produce exactly one row for each initializer row.
- Preserve the finding ID exactly.
- Use only
upheld,unclear, orrejected. - Keep explanations concise and evidence-based.
- Do not add commentary before or after the table in the file.
- Do not return the table inline in chat.
More from henryqw/skills
gh-autopilot
Run a standalone autonomous GitHub Copilot pull request review loop with explicit stage entry and event logs. Use when Codex should start from a user-selected stage (create PR, monitor review, or address existing comments), execute deterministic cycle transitions, and continue looping until Copilot reports no comments or the configured Stage 2 max-wait limit is reached.
10gh-pr-creation
Create a new GitHub pull request end-to-end when the user asks to open or create a PR. Use when Codex must turn local uncommitted work into a reviewable PR by making multiple scoped commits, running and passing all repository quality gates, renaming the branch so it reflects the changes, creating a Conventional Commits PR title, writing a PR description with summary/rationale/migration steps, and assigning Copilot as reviewer.
6gh-address-copilot-review
Handle GitHub PR review comments when comments are provided by the user as context. Use when Codex must evaluate comments one by one, classify each as actionable or non-actionable or needs clarification, implement only necessary fixes, keep changes scoped per comment, run validation, avoid intermediate pushes, perform one final push for the full batch, resolve addressed threads, respond to rejected comments with rationale, and re-request Copilot reviewer exactly once at the end via gh-assign-copilot-reviewer.
5triangulate
Evaluate supplied artifacts and return a consolidated findings table with evidence-based conclusions. Use this skill when the user wants a proposal, plan, code change, document, prompt, transcript, or other material reviewed through a structured multi-perspective evaluation instead of a single opinion.
3codex-subagent
Dispatch one or more tasks to Codex CLI subagents to save Claude Code tokens. Accepts explicit task descriptions, auto-selects sandbox (read-only vs workspace-write) and reasoning effort (high vs xhigh) based on task type, and collects structured results with durable artifacts.
2trueflow
Run the full generic trueflow pipeline by invoking `trueflow_initializer`, `trueflow_adversary`, and `trueflow_referee` in sequence, persisting stage outputs under `.context/trueflow/`, and returning a consolidated `findings.md` table. Use this skill whenever the user asks to "use trueflow" or wants multiple agents to review artifacts, solution proposals, coding implementation plans, documents, prompts, or other material and return adjudicated findings rather than a single opinion.
1