Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add AkshitIreddy/agent-skills --skill adversarial-refinement-loopgit clone --depth 1 https://github.com/AkshitIreddy/agent-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/akshitireddy/agent-skills/adversarial-refinement-loop)<a href="https://agentmods.dev/skills/akshitireddy/agent-skills/adversarial-refinement-loop"><img src="https://agentmods.dev/badge/skills/akshitireddy/agent-skills/adversarial-refinement-loop.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00043 | $0.01277 |
| Opus 5 | $0.00022 | $0.00639 |
| Sonnet 5 | $0.00009 | $0.00255 |
| Haiku 4.5 | $0.00004 | $0.00128 |
Grade A, and why
adversarial-refinement-loop scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 6d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 110 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Adversarial Refinement Loop
Treat refinement as an open-ended search, not a checklist that converges after one green build. Once invoked, keep starting new passes until the user explicitly pauses or stops, a genuine external blocker prevents useful work, or the host ends the run. A pass with no high-confidence defects is evidence to change the review lens; it is not a completion condition.
An ordinary response boundary, context compaction, temporary tool failure, exhausted lens, or clean result is a checkpoint, not completion. When the host supports persistent goals or automatic continuation, use it. If a hard host stop is unavoidable, label the run incomplete, preserve the exact ledger and next lens, and resume on the next supported continuation. A blocker ends the active run only when no safe, useful pass remains under any lens after reasonable alternatives have been tried.
Away-mode context
Users commonly invoke this loop immediately before going to sleep or becoming busy with something else. Treat their expected silence as permission to keep working within the already authorized scope, not as a request for frequent decisions or a signal to stop. Prefer evidence-backed, reversible assumptions; move to another useful lens when one lane genuinely needs user input. Absence does not expand permission to publish, push, release, contact people, delete data, expose private material, or make other external mutations.
If the user returns and says they are awake, available again, or wants to stop the loop, exit persistent loop mode promptly. Stop scheduling automatic new passes, preserve the current ledger and active findings, and continue any remaining requested work under ordinary collaboration rules unless the user also asks to pause or stop the project itself. Leaving loop mode is not the same as discarding the backlog.
Pass loop
Maintain a lightweight pass ledger containing the lens, evidence, changes, verification, and unresolved hypotheses. For every pass:
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 6d ago First seen · 110 lines · 43 tokens per session scan A 4a8e8868afea
adversarial-refinement-loop is a skill published in the GitHub repository AkshitIreddy/agent-skills (1 stars, last pushed 9d ago), licensed MIT. It adds 43 tokens to every session and 1,277 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
simplify
This skill should be used when the user asks to "simplify", "clean up the diff", "run simplify", "simplify the changes", "review changed code for cleanup", explicitly invokes "/simplify", or asks to "commit without simplify", "skip simplify for this commit", or "commit this but skip simplify". Reviews changed code…
resolve-pr-comments
This skill should be used when user asks to "address PR comments", "resolve PR feedback", "handle review comments", "fix PR issues", "respond to PR review", or explicitly invokes "resolve-pr-comments".
review-pr
This skill should be used when user asks to "review a PR", "review pull request", "review this pr", "code review this PR", "check PR.
signal-review
Review code in SIGNAL's one-line format with required severity per issue. Output is typed, structured, and chainable — not prose paragraphs. Use when user types /signal-review, "review this", "review my code", "check this PR", "review this file", or asks for a code review on any file, diff, or PR. Works standalone or…
adr-review
Review an ADR the user has written in docs/adr/. Use when the user says an ADR is ready for review or asks for feedback on a decision record. Never use this to write or rewrite an ADR for them.
chinese-code-review
Chinese-language guidance for giving professional, tactful code-review feedback while still calling out real problems.