What we know about specifications
What we have learned about specs, packaged guidance, and what actually helps an agent implement.
The clearest conclusion so far
A real specification gives an agent a better starting opportunity than leaving important behavior implicit. Making product decisions explicit reduces the consequential policy an agent must infer or invent.
With SeedSpec, you write specs as you normally would, but you are encouraged to package extra context modules with them—skills, evals, behaviors, and the rest. The result is one portable artifact you can version, check, and hand around as a .seedspec file.
Packaging those pieces together is what keeps implementations pointed at the same intent without forcing a specific stack or output.
What the current work indicates
Good Markdown is a strong baseline
A well-written Markdown spec is already a strong starting point for an agent. Extra structure, a manifest, or a valid package does not automatically improve the intent.
Specialized guidance can carry missing knowledge
Most agents will not already have the domain or technical knowledge a job needs. Skills and other packaged guidance can travel with the spec so that knowledge is there at implementation time.
Explicit intent reduces consequential invention
When important product decisions stay implicit, agents fill the gaps. Spelling those decisions out leaves less room for invented policy, states, or behavior.
Semantic authoring makes ownership clearer
When product intent, adopter choices, success criteria, and implementation guidance are easy to tell apart, people and agents can inspect and adapt the package without mixing those roles.
Skills can help, but they do not replace judgment
A relevant skill can improve what an agent produces, especially when the model is missing procedure or domain knowledge. It does not make the model reliable on its own, and consulting a skill is not the same as independently checking the result.