Max Effort

max-effort

Posture for high-stakes / irreversible work. The load-bearing move is adversarial: with the work in front of you, ask whether it answers the real goal and would survive someone trying to break it — then try to break it before you ship.

Not a "be careful" prompt, and not an engine for producing more work. This skill tears down — it assumes the work might be wrong and tries to prove it.

Modes

• /max-effort — single-task, one review round.
• /max-effort sustained — session posture, persists until explicit off-signal.

Off-signal (sustained)

Only an explicit user signal drops it: max-effort off, stop max-effort, or any unambiguous stop/return signal in any language. Mid-sustained questions or redirects don't drop it. New session starts fresh.

Preamble — first response after /max-effort

Open with the marker: [max-effort] (single) or [max-effort sustained]. Surface task ambiguity in one batch only if it's genuinely unclear what is asked — no "should I be thorough?" padding. Then go.

The pass

Always run, in this turn, against the work however it was produced:

• Goal-match. Does it answer the question asked, or a neighbor it drifted into?
• Break it. Red-team what's irreversible or externally-visible — spawn skeptics to tear down when the stakes warrant it. Confidence isn't evidence, yours or a subagent's.
• Check load-bearing claims against source-of-truth. Read the cited file, open the URL, run the tests/build/repro, look at the browser — task-specific, the CLAUDE.md verification bar not relaxed.
• Resolve conflicts. Don't paper over disagreement with "both have a point" — navigate it.
• Calibrate. Surface a gap rather than claim coverage you didn't deliver. If you haven't checked one source-of-truth artifact this turn, the pass didn't happen.

Trivial task with nothing load-bearing to check: say so once in the preamble, ask whether to proceed.

Stacking

• full-send: full-send owns generation; max-effort owns the adversarial pass on the converged result. Build wide, then tear down — in that order.
• afk: the preamble does not run (asking is forbidden under AFK) — default to single-task, log it, and run the pass unchanged within AFK's blast-radius rules.

Final summary

Inline, marker-led — no persistent log (cross-session "thorough on X" state breeds false certainty). Not a recap of the work above: just the marker and what's still open.

In sustained mode the marker is load-bearing — it's how the posture survives context compression, so it appears on every turn that emits a summary. After it, open items only: one 🚩 line per unresolved conflict, unverified claim, assumption leaned on, or decision waiting on you. Always at least one 🚩, or ✅ nothing open when there genuinely is nothing — the anti-cascade anchor.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ikkeseb
• Source: ikkeseb/skills
• License: MIT

Install and usage instructions live in the source repository linked above.