Neural plan
Synthesize approved discovery into a stable, shareable product spec with public contracts and testing decisionsFrom its SKILL.md
npx -y skills add juancruzrossi/neural --skill neural-planAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.9 KB, as published. Nobody here has run it
Neural Plan
Turn approved discovery into .neural/wip/<feature>/PLAN.md: the stable product
specification for teammates and future agent sessions. It defines what will be
built, not an implementation checklist. Do not create a separate PRD.
Establish the contract
- Parse
$ARGUMENTS. When--skillsis provided, keep the requested skills in mind and load each one on demand when relevant to the current work. - Resolve the feature from the remaining selector or
.neural/wip/. If none exists, point to neural-interview; if several exist, ask which one. - Require
CONTEXT.md. Read it, every feature ADR, and.neural/knowledge/when present. - Trace the interview lenses into the specification without duplicating rules: affected party into Problem; done into Product Outcome, behaviors, and acceptance; non-negotiable priorities into the relevant behaviors, decisions, and testing; proposed proof into Testing Strategy. Return any lens that cannot be represented without guessing to neural-interview.
- Inspect related code and tests to understand current behavior, domain language, existing public interfaces, and testing precedent. Use that knowledge to keep the spec realistic, not to predict an edit list.
- Normalize each public operation and lifecycle transition into its input, success output, errors, next state, and observable side effects. If repeated statements conflict, return the exact conflict to neural-interview.
- Run a counterexample check: imagine two reasonable implementations that satisfy the written context. If callers could observe different results for the same input or lifecycle, return the smallest distinguishing product decision to neural-interview.
Do not reopen resolved decisions. Implementation freedom is valid only inside the approved decision boundaries.
Write the specification
Use PLAN-FORMAT.md. Make it understandable without the interview transcript: state the problem, proposed product outcome, observable behaviors, public contracts, consequential implementation decisions, testing strategy, acceptance criteria, and exclusions.
PLAN.md becomes the canonical approved specification. Link CONTEXT.md and
ADRs for discovery history and rationale rather than copying the conversation.
When --skills was provided, include ## Skills to load as a reminder of
the requested skills. Load them on demand when relevant. Omit the section
otherwise.
Keep decisions at product, public-interface, module, schema, or architectural level. Do not include exact file paths, an exhaustive task breakdown, edit order, progress tracking, or code snippets likely to become stale. A compact prototype fragment is valid only when it expresses an approved contract more precisely than prose. Write every code fragment in Python.
Prefer an existing public interface. When the feature creates or materially changes one, load INTERFACE-DESIGN.md and compare alternatives before recording the decision.
Testing decisions must name and observe the public interface used by callers or another established public interface. Internal database, private method, filesystem, or log inspection cannot prove product behavior unless that surface is itself part of the public contract.
Assign each behavior an evidence mode. Use outcome by default. Use
test-first only for a bug reproduction, a pre-approved executable example,
or an explicit user decision, and record the reason. Include a representative
scale scenario only when input growth can affect the promised outcome.
Define each behavior once with an ID. Testing decisions and acceptance criteria refer to those IDs instead of repeating the behavior. Every behavior ID must appear in both sections; no specified behavior sits outside the release gate. Include failure states, ordering, retries, concurrency, or rollback only where the product promises them.
Planning is complete only when a teammate can explain what is being built and why, behavior IDs have complete testing and acceptance coverage, every behavior has a public way to observe it, no product ambiguity remains, and the document contains no implementation checklist or placeholder.
Optional adversarial review
If another supported coding agent is installed, offer one cross-review. Load CROSS-REVIEW.md only if the user accepts. Never change the specification from external feedback without explicit approval. When declined, omit review output.
Report the behavior count and major public-interface decisions, then suggest:
Ready to execute? Run neural-execute.
Leave the specification local. Never stage, commit, or push; preserve pre-existing staged and unrelated changes as found.