agentleFS
dgrissen2dgrissen2/persona-review-kit/skills/claude-strategy-review/SKILL.md

claude-strategy-review

Skill0 starsChanged 3 months agoMITView on GitHub

Run an independent Claude Opus 5 adversarial strategy and logic review with tools enabled and optional persona lenses. Use for strategy docs, research conclusions, optimizer proposals, risk frameworks, decision memos, plans, or arguments that need stress-testing; generic by default and not trading-specific.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 4 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this to stress-test a strategy, argument, proposal, research conclusion, risk framework, optimizer design, decision memo, or any other …

    line 14
  2. Distinctive:

    If the user provides personas, use them.

    line 36
  3. Distinctive:

    Run a generic-only review:

    line 49
  4. Distinctive:

    Run with persona aliases or persona file paths:

    line 57
  5. Very specific:

    If Claude times out or the wrapper cannot complete, report TOOLING FAILURE with the artifact path.

    line 83
  6. Very specific:

    Default permission mode: auto

    line 107
  7. Very specific:

    Default heartbeat: 20 seconds

    line 110
  8. Very specific:

    Default timeout: dynamic — soft budget auto-estimated from prompt + document sizes and effort, deadline auto-extends while streaming, stall …

    line 111

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Claude Strategy Review

Independent adversarial logic review using Claude Code, pinned to claude-opus-5 with --effort xhigh, --tools default, and --permission-mode auto.

Focus

Use this to stress-test a strategy, argument, proposal, research conclusion, risk framework, optimizer design, decision memo, or any other plan where the question is "does this logic hold up?"

The generic review checks:

  • conclusions that do not follow from premises
  • hidden, circular, fragile, or unstated assumptions
  • made-up thresholds, constants, weights, or scoring rules
  • missing evidence, counterexamples, and falsification tests
  • incorrect math, units, base rates, or time horizons
  • feasibility, incentives, adoption, and operational constraints
  • downside, tail risks, failure modes, monitoring, and reversibility
  • contradictions between goals, constraints, and proposed mechanisms

Personas

Personas are optional lenses, not a fixed enum. The runner discovers persona markdown files through the shared persona registry. Examples include quant, pm, brent, architect, red-team, marcus, or a direct persona file.

If the user provides personas, use them. If the user asks for a panel or persona lenses without naming them, ask one concise question about which lenses to include. Otherwise run the generic logic review by default.

List discovered personas with:

python3 ~/.codex/skills/claude-strategy-review/scripts/claude_strategy_review.py \
  --list-personas

Workflow

Run a generic-only review:

python3 ~/.codex/skills/claude-strategy-review/scripts/claude_strategy_review.py \
  <strategy_file> \
  [context_file ...]

Run with persona aliases or persona file paths:

python3 ~/.codex/skills/claude-strategy-review/scripts/claude_strategy_review.py \
  <strategy_file> \
  [context_file ...] \
  --personas quant,pm,/path/to/custom_persona.md

Each persona pass is independent: do not feed one persona's output into another. The final synthesis may see all outputs and should deduplicate agreement, disagreement, and unique catches.

The runner streams Claude progress by default, emits heartbeat updates to stderr, and writes durable artifacts under ~/.cache/agent-review-runs/. Each Claude attempt records prompt.txt, stdout.stream, stderr.stream, events.jsonl, result.md, and timeout diagnostics. On SIGINT/SIGTERM/timeout, the wrapper terminates only its own Claude process group.

Timeouts are dynamic: the soft timeout is auto-estimated upfront from the prompt plus strategy/context document sizes and effort, then the deadline keeps extending while Claude is still streaming output, bounded by a stall watchdog and a hard cap. Heartbeats, partial-output progress, and deadline-extension notices on stderr are the liveness signal — do not impose a shorter outer shell timeout or kill the process on brief silence; the wrapper always exits on its own.

If Claude times out or the wrapper cannot complete, report TOOLING FAILURE with the artifact path. A tooling failure is not a strategy verdict and must not be treated as approval, rejection, or evidence about the strategy itself.

Canonical personas live at ~/.config/persona-review-kit/personas/, and the shared resolver lives at ~/.codex/skills/persona-registry/scripts/persona_registry.py.

Output Contract

The script returns a generic review, optional persona reviews, and optional synthesis:

## Claude Strategy Review — Generic Logic
## Claude Strategy Review — <Persona Name>
## Claude Strategy Review — Panel Synthesis

Notes

  • Default model: claude-opus-5
  • Default effort: xhigh
  • Allowed effort values: low, medium, high, max, xhigh
  • Default tools: default
  • Default permission mode: auto
  • Allowed permission modes: auto, default, dontAsk, plan; edit-accepting and permission-bypass modes are intentionally rejected for review runs over untrusted strategy/context
  • Exit code 0 means the wrapper completed and emitted a strategy review; callers must inspect the review text. Exit code 1 means wrapper/tooling failure or invalid invocation.
  • Default heartbeat: 20 seconds
  • Default timeout: dynamic — soft budget auto-estimated from prompt + document sizes and effort, deadline auto-extends while streaming, stall watchdog 300s (CLAUDE_REVIEW_STALL_SEC), hard cap 3600s (CLAUDE_REVIEW_MAX_TIMEOUT_SEC). Pass --timeout-sec N (or CLAUDE_STRATEGY_REVIEW_TIMEOUT_SEC) to pin the soft budget instead.
  • Troubleshooting flags: --no-stream, --heartbeat-sec N, --safe-mode, --disable-slash-commands
  • Override with CLAUDE_STRATEGY_REVIEW_MODEL, CLAUDE_STRATEGY_REVIEW_EFFORT, CLAUDE_STRATEGY_REVIEW_TOOLS, CLAUDE_STRATEGY_REVIEW_PERMISSION_MODE, CLAUDE_STRATEGY_REVIEW_TIMEOUT_SEC, CLAUDE_STRATEGY_REVIEW_HEARTBEAT_SEC, or the shared runner watchdog vars CLAUDE_REVIEW_STALL_SEC / CLAUDE_REVIEW_MAX_TIMEOUT_SEC.

Intent and focus (automatic)

Derive these from the user's request — they should not need to type flags. From how they phrase the ask, infer intent (what the work is trying to do) and focus (what to weight) and pass them as --intent / --focus. For example, "review the auth refactor on this branch, focus on whether the logic is right and skip style" -> --intent "auth refactor" --focus "logic correctness, skip style".

The text is appended to the prompt as author framing, never as another reviewer's findings (the review stays independent). A guardrail still surfaces every CRITICAL/HIGH finding even when it falls outside the focus.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.