Tcuzzo/backs-aios-skills

---

name: absorb description: Use when you need a capability an open-source project already provides — adopt it and re-engineer it as a native skill instead of inventing a duplicate. Trigger words: absorb, adopt, port, re-engineer, ingest a repo, prior art, capability port, make this native. license: MIT

Absorb — Adopt Prior Art, Don't Reinvent It

Effort: light — a repo ingest (metadata probe + shallow clone) and a cross-family grade of the port. Removes: reinventing a capability prior art already solved — the from-scratch duplicate you'd owe every bug for.

Capability is king. A repo is a vehicle for a capability. When you need something an existing project already does, do not build a duplicate from scratch, and do not clone-and-paste. Find the best prior art, extract the capability, re-engineer it to fit your harness, and cite the scaffold. The citation is a fact, not a decoration.

When to run

• You are asked to add a capability (a tool, skill, agent, pipeline) that open source likely already solves.
• You are about to git clone and copy code verbatim — stop; this is the path instead.
• Skip it for a single snippet, a config value, or a fact lookup. Just read those.

Steps

1. Hunt prior art first. Search before you build. A duplicate you invent is worse than a scaffold you adopt: you inherit zero field testing and you owe all the bugs.
2. Ingest beyond the README. Pull the project metadata (license, activity, language) from the platform API. Shallow-clone into a scratch directory. Read the code and the tests. The README is marketing; the code is the truth.
3. Run the trust gates.
   • License: permissive (MIT / Apache / BSD / MPL) = safe to re-engineer. Copyleft (GPL / AGPL) = technique-only — re-engineer the idea, never copy the code. No license = treat as all-rights-reserved, technique-only. Non-commercial terms = a blocker; take it to your human.
   • Shady scan: grep for cloak / spam / fake-review / scam patterns. Flag loudly.
   • No wild installs: never pip install / npm install an unvetted dependency (typo-squatting is a real supply-chain attack). Re-engineer as thin code over your own primitives instead.
   • Is the capability real? Verify claims against independent evidence. A seller's blog is a claim, not evidence. Verdict: real / hype / scam / unverifiable.
   • Bounded egress: anything the adopted version fetches must be throttled, cached, and killable.
4. Deconstruct into a capability map. For every ability the project provides, record: what it does, how, its load-bearing seams, its bloat or risk, what you can reuse from your own stack, and whether it lands native or behind a thin adapter. Every capability is preserved or refuted with evidence. A silently dropped capability is a defect.
5. Write the re-engineering spec. The seams to build, the bloat you are dropping (recorded loudly, never silently), and one failing contract test per capability that asserts a real side-effect — a file, a database row, real output. Mock only a paid external API's transport, never the logic.
6. Reconstruct red-first. Commit the failing tests, then build until green across the whole seam. A model from a different family than the builder grades the result — the builder never grades its own work.
7. Cite and record. Write the scaffold credit where the capability now lives: author, project, license, what is borrowed (the scaffold) and what is yours (the re-engineering). Never invent a credit. Never strip one.

Hard rules — any one of these fails the skill

• Copying code verbatim instead of re-engineering the capability.
• Building a duplicate without ever searching for prior art.
• Trusting the README or a marketing page over the code.
• Installing a wild dependency instead of re-engineering the technique.
• Copying copyleft or unlicensed code (technique-only, always).
• Dropping a capability without a written refutation.
• Mock theater in a capability test — the test must touch a real side-effect.
• Shipping without the scaffold citation.

Works well with

• red-first — the contract tests that guard each capability.
• sniper-testing — real side-effects, no mock theater.
• blind-tribunal — cross-family grading of the port.
• decision-bar — license blockers and taste calls go to your human; everything else executes.