Challenge Project

A battle-tested methodology for high-difficulty AI-agent projects — distilled from real production practice. Scale by complexity, verify by evidence, evolve the harness.

[English] | 简体中文

---

Table of Contents

• Why
• Quick Start
• Core Mechanisms
• Versions
• Platform Adaptation
• Credits
• Contributing
• License

Why

• 能力在装具，不在模型 — Agent = Model + Harness; the main lever is the harness (rules, checklists, prompts, flows), not the model.
• loop on evidence, not confidence — terminate with tests/evidence, never "looks right".
• 该轻则轻，该重则重 — L0-L3 grading keeps token efficiency optimal.
• DeepSeek Harness efficiency — grade every task (L0-L3) so tokens go where they matter; terminate on evidence instead of guesswork; break retry loops (≤2 attempts) before they burn context; evolve rules after each delivery, so recurring work costs a fraction (up to 6-10x fewer tokens).

v2.0 fuses three frontier papers on harness engineering:

1. Harness Engineering (David Daniel, 2026-06) → three separation modes (planner/executor, writer/reviewer, initializer/coder)
2. The Last Harness You'll Ever Build (arXiv:2604.21003) → Agent = Model + Harness, Harness Evolution Loop
3. The Harness Layer as Control, Agency, and Runtime → CAR decomposition + HarnessCard

Quick Start

1. Drop it into your persistent instructions: copy openclaw/SKILL.md (OpenClaw) or dsh/SKILL.md (DSH) into your instruction file / skills directory (CLAUDE.md for Claude Code, AGENTS.md for Codex/DSH; see the adaptation guide for others).
2. Run one L1 task as a drill: grade it (one-line L0-L3), then follow the corresponding flow.
3. Self-check: walk the「自检清单」at the end of the SKILL after the task; run the「复盘模板」for L2+ tasks.
4. Go cross-platform: docs/adaptation-guide.md maps every mechanism to OpenClaw / DSH / Codex / Claude Code / Cursor / PI primitives.

Core Mechanisms

#	Mechanism	One-liner
1	HARNESS-LOOP-GRAPH	HARNESS = boundary (the rig), LOOP = evidence-driven feedback, GRAPH = flow topology; self-similar at every scale
2	L0-L3 Grading Gates	trivial / light / standard / major — scale the process and token budget by complexity; irreversible, paid or multi-step work auto-upgrades
3	Three-Stage Pipeline	research (design the HARNESS) → implement (run the LOOP) → accept (triple review + diversified verification signals)
4	Decision Council	triad/standard multi-perspective adversarial review; verdict with evidence labels and Kill Criteria; honest escalation, never forced consensus
5	Harness Evolution Loop	post-delivery retro → adversarial diagnosis of failure modes → update the rig (rules/checklists/flows) → human-approved effect
6	HarnessCard	when reporting capability, report the harness layer too (Control / Agency / Runtime / verification signals / failure modes)

Grading gate at a glance

📜 Verdict · [Project Name]
【Grade】 L0 trivial / L1 light / L2 standard / L3 major
【Verdict】 ✅ pass / ⚠️ conditional / ❌ reject
【Flow】 direct / light / three-stage / full
【Rationale】 …

Versions

Path	Description
openclaw/SKILL.md	OpenClaw version v2.0 — the full methodology
dsh/SKILL.md	DSH adaptation — core mechanisms kept, OpenClaw-specific dependencies mapped to DSH equivalents
docs/adaptation-guide.md	Platform mapping: OpenClaw / DSH / Codex / Claude Code / Cursor / PI

Platform Adaptation

See docs/adaptation-guide.md — a mechanism × platform mapping table (rules → instruction files, planning → todo/plan primitives, review → subagents/parallel sessions, checkpoints → files/goals, cost → real billing APIs), plus a PI adaptation section and a no-subagent fallback FAQ.

Credits

Distilled and refined by multiple AI collaborators and a human partner through real project practice — deliberately anonymous. Maintained and released by hlxstc (2026-08-15). See CREDITS.md.

Contributing

Feedback from practice is the lifeblood of this methodology. See CONTRIBUTING.md — issues, PRs and platform adaptation experiences are all welcome.

License

MIT — free to use, modify and distribute with attribution.