Extend Before Inventing

Extend Before Inventing

The most expensive code in a project is the second implementation of something it already had. Two versions of one idea drift apart, each collects its own bugs, and every future reader pays to learn which one is real. Observed across production agent runs, agents invent parallel systems not from need but from search failure: creating felt faster than finding. This skill makes finding the first move, every time.

1. Search first, and mean it

Before creating anything, search the codebase for the thing you are about to build: the concept, the likely names, the sibling that must have solved this already. Read what you find. A close match gets extended (a parameter, an option, a new case) rather than re-implemented next to itself. "Close but not exact" is the design prompt, not the permit to fork: shaping the existing thing to cover your case is the work.

2. The precedence order: built-in beats library beats custom

When nothing in the repo covers it, prefer the platform's built-in, then an established library already in the dependency tree, then a new dependency, and only last a custom implementation. Before wrapping or replacing a primitive, prove it is insufficient: state what you need, show the primitive cannot do it, and record that. An abstraction justified by "cleaner" rather than by a demonstrated gap is a liability with good manners.

3. New content is data on an existing spine

Most "we need a new system" moments are actually "we have new content". Adding a template, a rule, an entry, a page, or a case should mean adding data to a structure that already renders or executes it, not building a second structure. If adding one more item requires new machinery, fix the spine so it takes items as data; do not clone the spine.

4. The invention cap: one sanctioned new system per work cycle

Genuinely new systems are sometimes right. Budget them: at most one per work cycle (a session, a sprint, a commission), chosen deliberately and named as such, with the search that came up empty recorded. The cap is not against invention; it is against invention as a reflex. If a second new system feels necessary in the same cycle, that is the signal to stop and re-examine the first.

5. Shorter is the success metric

After a change that touches existing structure, the file should end shorter or no longer than it started, unless the growth is justified in writing. Deletion of a duplicate path counts as progress and belongs in the report with the same pride as a feature: every removed parallel implementation is a permanent drop in the cost of understanding the project.

Anti-patterns

• Creating because searching felt slow. The five minutes saved becomes a permanent tax.
• Forking a near-match instead of extending it. Two copies never converge again.
• Wrapping a primitive you have not proven insufficient. The wrapper hides the

primitive's real API and adds a surface nobody asked for.

• Building machinery to hold one item. One item is content; put it on the spine.
• A "quick parallel version just for now". There is no such thing; parallel versions

are permanent the moment a second caller appears.

Output

Return, for each new artifact proposed or reviewed:

• The search performed and what it found (or that it found nothing).
• The decision: extended existing / built-in / existing library / new dependency / custom, with the insufficiency proof where required.
• Whether the invention cap for this cycle is already spent, and by what.
• Net line delta for touched files, with justification if anything grew.
• Any duplicate path deleted, called out as progress.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ariaxhan
• Source: ariaxhan/the-agent-library
• License: MIT

Install and usage instructions live in the source repository linked above.