Your model is not the bottleneck.

Your harness is.

if your agent only works when AGENTS.md keeps growing: you do not have alignment you have prompt debt fix: smaller root router harder verification gates repo-specific evals

• ClawPond landing page: https://clawpond.com/harness-engineering
• Skill detail page: https://clawpond.com/skills/harness-engineering
• GitHub repository: https://github.com/LeoFanKm/harness-engineering
• Install command: claw install harness-engineering

• Skill entry: SKILL.md
• Repo audit script: scripts/harness_audit.py
• Eval scaffolding: scripts/scaffold_repo_eval.py
• Eval run scaffold: scripts/scaffold_eval_run.py
• Eval aggregation: scripts/aggregate_repo_eval.py
• Prompt kit: references/prompt-kit.md
• Eval playbook: references/eval-playbook.md
• Scoring rubric: references/scoring-rubric.md
• Research map: references/research-map.md
• Landing copy: docs/landing-copy.md
• Release notes: docs/release-notes.md

Most agent failure still comes from the same four boring causes:

• the repo is illegible
• verification is weak
• old paths still grow
• autonomy exceeds the fences

This skill packages the counter-pattern into something a repo can actually enforce:

1. shrink the root instruction surface into a router
2. pin durable truth into named docs and commands
3. fence bad paths with scripts, CI, and compatibility boundaries
4. prove the change with meaningful verification and repo-specific evals

• Root context shrank from 229 lines to 66
• Historical benchmark: 92.0 review-loop vs 47.3 single-pass
• Release chain now guards skill data, landing pages, schema, sitemap, and smoke coverage together

1. Read SKILL.md.
2. Run python scripts/harness_audit.py <repo-root>.
3. Use references/prompt-kit.md to choose the smallest intervention.
4. Use references/eval-playbook.md when you need repo-specific evals.

• WeChat group: scan the QR code below for the OpenClaw / Harness Engineering community.