Absorb
Intent
- Goal: Produce a verified, contrasted adoption shortlist of external AI-harness practices for this repository.
- Limits:
- Analyze harness and agent configuration, not the external project's business code quality.
- Clone or update a git URL only with user awareness and to a clearly stated local destination.
- Never auto-adopt, auto-commit, or silently edit this repository as a result of the audit.
- Fails if:
- A README claim is promoted as proven behavior without source, activation-point, or wiring evidence.
- A practice is recommended without contrast against this repository.
- Findings are fabricated when a red flag applies: the path does not exist, the user wants general code review, no AI-config surface exists, or practices are documented but not wired. Surface these explicitly.
Context
- Reads:
- Targets: git URL(s) and/or local path(s); optional focus (skills, agents, prompts, workflows, installer patterns, comparison themes). When no git URL or local path is given, ask for at least one before starting; treat remaining arguments as the analysis focus.
- External AI entrypoints:
AGENTS.md,CLAUDE.md, command, agent, and skill folders, config files, prompt files, AI plugins. A repo is a valid harness even withoutAGENTS.md,agents/, orskills/; its AI surface may live in plugins, runtime config handlers, and maintenance scripts. - Repo-native contrast:
AGENTS.md,.ai/atl/skill-registry.md,domains/*/README.md,domains/*/skills/*/SKILL.md,skills/*/SKILL.md, relevant local artifacts.
- Never reads: memory as a substitute for source evidence; delegate read-heavy discovery to subagents or researcher passes when available.
- Persona: skeptical harness auditor that avoids cargo-cult adoption.
- Scope: the user wants to learn from an external repository's or local project's AI-harness setup: agents, skills, prompts, orchestration rules, or instruction patterns. Not for application code review, bug hunting, product architecture review, or test-quality audits.
Expectations
- Scenarios:
- WHEN targets are given THEN they are resolved (local paths verified) and the list confirmed before deep analysis.
- WHEN this repository already covers a practice well THEN it is marked
no adoption needed. - WHEN the user explicitly asks not to persist THEN the file is skipped and the same structure returns as a concise chat summary.
- Done means: the report exists at
.ai/absorb/YYYY-MM-DD-<slug>.md(<slug>is short kebab-case focus or target; fallbackexternal-practices; create.ai/absorb/if missing), unless chat-only was requested. Return: resolved targets; key verified practices; rejected or conditional practices with reasons; our advantages; shortlist with proposed target paths in this repository; report path (always, unless the user opted out of persistence). - Quality needs: each recommended practice has file-level evidence; at least one
no adoption neededorOur Advantagesitem; no unverified README-only claims in the shortlist; no commit or automatic adoption.
Method
Map each harness. Record how each candidate practice is triggered or wired, and per practice:
Field What to record Practice Clear name Mechanism File path and concrete evidence Status Verified / partial / aspirational Strength Why it helps Weakness Limitation or trade-off Contrast. Carry forward only genuinely better or meaningfully missing practices. Record at least three Our Advantages points.
Adversarial filter. Is it truly verified? Better, or just different? Does it fit this repo's OpenCode-only, compact, contract-focused conventions? Is adoption cost proportional to benefit? Keep only
ADOPTorCONDITIONALitems.Report sections: sources analyzed; one section per external project; verified strengths; Our Advantages; ranked adoption shortlist with target file paths in this repository; immediate text-only wins vs. larger follow-up work.
Attribution
Created by abdi. Adapted for this repository by andresnator.