Develop
Personal agent skills for Claude Code, Cursor, and other AI coding agents.
npx -y skills add guillevc/skills --skill developAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 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.
What its author says it does
Copied from the file, not written here
Take a unit of work from fuzzy idea to shipped, on one loop. Interrogate the plan, preview the change as blast radius plus sample code at the confirm gate, build against the spec, freeze the decisions the code proved, verify, optionally commit or open a PR. Stop at the confirm gate to think without building. Never merges. User-invoked.
SKILL.md
4.9 KB, as published. Nobody here has run it
Develop
One loop from idea to shipped, against the living spec (glossary + ADRs). The spec trails the code: a decision earns an immutable ADR only after code proves it; until then it's a mutable draft in the brief, never in the ADR directory.
Gates
- Autonomous: ask, brief, draft decisions, run checks, assemble the PR.
- Hard: the confirm gate before any code change; the draft review before freezing; durable writes
go through
record; every commit, defaulting to one commit on the current branch, with branch, push, and PR only on request. Never merge.
The loop
Spine: discover → confirm gate → build → freeze → verify → deliver. The confirm gate is a hard stop: never write code until the user gives an explicit go-ahead. Re-enter discover from build or review whenever your mind changes.
- Discover. Map the plan into a tree of named branches, then walk it branch by branch, resolving
each decision and its dependencies before the next. Interview the user until you both share one
understanding of every branch; loop rather than settle for a single pass. Draft the brief only then.
- Interrogate one branch at a time: ask the single sharpest question that cuts the most uncertainty, with your recommended answer. Ask one question, then wait for the answer before the next. Batching questions bewilders the user and defeats the interview. If the spec or code already answers, read it instead of asking.
- Sharpen vocabulary: challenge a term that conflicts with the glossary, pin a fuzzy or overloaded one to a single canonical term, stress-test relationships with concrete edge-case scenarios, and when the user asserts how something works, check the code agrees.
- Brief: one resolution per branch, acceptance-shaped. It constrains the build without being written. Record nothing pre-code unless the human asks (a fact something outside this loop needs now, or insurance against context loss).
- Confirm gate (hard stop). Present the brief and a build plan, then wait. Write no code
until the user gives an explicit go-ahead; no go-ahead, no build. Stopping here is a valid end:
pure discovery, nothing built, nothing frozen.
- Build plan: make the brief code-shaped so the user sees each decision before you write it. Blast radius — every file to create or change, one line each on why. Approach — a couple of sentences on how. Sample the decisions, not the plumbing — show code only for the hard-to-reverse, interface-shaping calls (the ones that would become draft decisions); leave reversible choices unwritten, since the code tests them in build. Don't sketch the whole diff; that's waterfall and double work.
- Build. Decide reversible choices on your own; the code tests them. A hard-to-reverse choice
becomes a draft decision in the brief: honor it as the current working decision when making
further choices, but keep it mutable. Rework it, never treat it as a ratified ADR, never write it
to
docs/decisions/yet. - Re-enter discover whenever building or review changes your mind, including when the user asks
for changes. Reworking a draft is an edit, not a supersede. Re-entry runs the rest of the spine
again: changed code goes back through freeze, so
auditre-runs on it. Audit is not one-shot. - Freeze (post-code). Run
audit. Present the full set of draft decisions and term changes the code proved, in one place, for the user to review before anything sets in stone. Fold survivors into the spec viarecord, which gates each ADR and glossary edit. An ADR stays soft until step 7 commits it: rework or drop it in place, no supersede. Trim behavioral prose that crept in. Surface pre-existing drift you didn't cause; don't fix it or grow scope. - Verify. Run the project's tests (
go test ./..., thepackage.jsonscript, the make/just target, else ask). Must pass. - Deliver. Default is a single commit to the current branch, and it waits for the user's explicit approval: present what you'd commit and wait. A local commit is a hard gate, not a reversible step; commit nothing without a go-ahead, outward-facing or not. Branch, push, or open a PR only on the user's explicit request; none is the default. Never merge.
Ad-hoc work skips develop; the guards (record, audit) fire during any work. The spec carries a
unit across sessions; un-frozen drafts don't.