v1.1.0 — RL14: Concede on the merits, never to please

A fourteenth redline. Behavioral change to every judging mode (drill, detective, and factual claims anywhere).

RL14 — "Concede on the merits, never to please"

Detective mode's only runtime safety net — the coach concedes when the user catches a flaw the answer key missed — previously lived only in modes/detective.md, below redline level, with no guard against the opposite failure: defending a self-authored key by authority (stubbornness). Because the same model writes the case, the key, and the frame the key is judged against, all three carry the same blind spot. RL14 lifts the net to floor level and makes it two-way:

• Never defend by authority. "The key says so" is not a defense; neither the key nor its frame is self-evidently right.
• Never concede to please. A concession not anchored to the specific step where the user's reasoning is established is fabricated — withheld like a false defense.
• Produced, not felt. Before ruling, the coach writes out a four-fact reconstruction (user's claim / key's claim / frame criterion + is-it-itself-sound / verdict reading off the first three). A wrong key gets a short admission, not a balanced-sounding tribunal.
• Translation discipline locks three pivot words that collapse in Chinese (改判 not 讓步; 成立 not 有道理; 理據 not 優點).

Honest residual limit

Prompt text cannot make reverse-fitted "theater" impossible — a model can fill the four slots in form while fitting them to the answer it wanted. The reconstruction raises the cost and leaves it on the page for the user's next turn to attack; the real backstop is that the user keeps the floor to challenge again. Defeating it outright would need a code harness (independent verifier) — out of scope for a prompt-only skill.

Review

Three rounds of cross-model review (codex gpt-5.5 + gemini 3.1-pro, converging independently) + security review (no finding). lint 299/299; canonical and claude-ai mirror byte-identical on the RL14 block; new RL14 invariants mutation-tested.

Assets

• critical-thinking-for-humans-claude-ai.zip — upload to claude.ai as a skill
• critical-thinking-for-humans-portable.md — single-file edition for any frontier model's chat window

v1.0.2 — practice-environment positioning + eval stability/validity split

[1.0.2] - 2026-06-14

Positioning and eval-honesty pass after a second round of cross-model critique.
Documentation only; no behavior change to the four modes.

• Positioning: a practice environment, not a transfer claim. The README no
  longer sells itself as a fix for carry-over. It is a place to practice the
  move; whether a named move transfers stays the field's open question, stated
  once rather than as a repeated hedge. The "rigor does not transfer on its own"
  motivation is kept (a true phenomenon, not a promise). Removed the
  "trained into its blind spot" harm framing: under a practice-environment
  framing a wrong key is a move to push back on, and the user judges their own
  level.
• Eval metric honesty: stability vs validity. Cross-model agreement is
  documented as a stability / reproducibility measure only (models share
  training data, so they share blind spots), never as correctness. Added
  protocol 1b — a human validity anchor: a person competent in the twelve
  structures cold-solves a frozen sample, human-agreement is the primary
  validity number, cross-model agreement demoted to a secondary stability
  statistic. No run yet, so no validity claim. New
  docs/evals/human-validity-anchor-TEMPLATE.md.
• SKILL.md description em-dash (root + claude-ai overlay): the one
  user-facing string the prior voice pass missed, changed to a colon. Trigger
  keywords untouched; routing unchanged.

v1.0.1 — research-grade hardening

[1.0.1] - 2026-06-14

Research-grade hardening after cross-model critique. No behavior change to the
four modes; this release sharpens honesty and adds manifest-validity CI.

• Manifest parse fix. SKILL.md (root and the claude-ai overlay) wraps its
  description in a >- block scalar so the embedded Triggers: colon is
  literal text. The prior unquoted value parsed as a second YAML mapping key and
  failed strict parsers and marketplace frontmatter validation.
• New scripts/check_manifests.py + CI (pre-build and post-build): parses
  every SKILL.md frontmatter, plugin.json, marketplace.json, and the
  SKILL.md inside the shipped zip. A missing root manifest, an absent zip
  member, or an unclosed frontmatter line now fails the gate.
• Honest positioning. The transfer claim is stated as the open hypothesis it
  is (grounded in the contested far-transfer literature and a moderate,
  heterogeneous meta-analytic effect), not a settled result. The README and
  modes/drill.md now state plainly that a drill key is written and self-audited
  by one model with no independent sign-off.
• Drill challenge window. A new Session Flow step: after the dissection the
  coach stops, invites a challenge to the key, and resolves it before the
  passport write, so a key it cannot defend on the merits never enters the
  longitudinal stats. Locked by the invariant lint.
• docs/evals/ evidence framework (templates only): the headline metric is
  cross-model agreement, not single-model self-scored accuracy. No empirical
  effectiveness claim until a result file exists.
• README prose normalized (em dashes removed); invariant-locked sentences kept
  verbatim.

v1.0.0 — first public release

A gym for your critical thinking — the AI is the coach, you do the work.

Four modes across distinct epistemic stances:

• drill — judge stance. 12 argument structures, single defensible answer, commit before analysis.
• scene — Socratic stance. 6 interpretive frames, never ranked, plus a fallacy-recognition track.
• expedition — guide stance. Audit verified impossible-tier problem packs (16, cross-domain).
• detective — guide-and-judge stance. Crack a runtime-generated multi-layer case flaw by flaw.

Plus a manipulation-recognition domain, thirteen redlines, and a local-only passport that tracks your longitudinal blind spots.

Install

/plugin marketplace add Imbad0202/critical-thinking-for-humans
/plugin install critical-thinking-for-humans

Or as a plain skill:

git clone https://github.com/Imbad0202/critical-thinking-for-humans ~/.claude/skills/critical-thinking-for-humans

No Claude Code? Grab critical-thinking-for-humans-portable.md below and paste it into any frontier model's chat. For claude.ai, upload critical-thinking-for-humans-claude-ai.zip.

Architecture

See docs/ARCHITECTURE.md for the full diagrammed map.

License: CC BY-NC 4.0.