develop skill
Develop end-to-end changes needing decisions, multiple files/projects, risk management, or delegable phases through research, optional innovation, spec, implementation, validation, independent review, and remediation.
- GitHub stars
- 6
- Files
- 3
- Last commit
- Oct 18, 2026
- License
- —
- Allowed tools
- —
What does the develop skill do?
develop is an agent skill. Develop end-to-end changes needing decisions, multiple files/projects, risk management, or delegable phases through research, optional innovation, spec, implementation, validation, independent review, and remediation.
How to install the develop skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the develop skill in Plugin
/plugin marketplace add leszekgruchala/buddy /plugin install buddy@buddy
Install the develop skill in Claude Code
~/.claude/skills/develop/SKILL.md
$ npx degit leszekgruchala/buddy/skills/develop ~/.claude/skills/developInstall the develop skill in Claude Code (project)
.claude/skills/develop/SKILL.md
$ npx degit leszekgruchala/buddy/skills/develop .claude/skills/developInstall the develop skill in Codex
~/.agents/skills/develop/SKILL.md
$ npx degit leszekgruchala/buddy/skills/develop ~/.agents/skills/developInstall the develop skill in Gemini CLI
~/.gemini/skills/develop/SKILL.md
$ npx degit leszekgruchala/buddy/skills/develop ~/.gemini/skills/developInstall the develop skill in Claude.ai
npx degit leszekgruchala/buddy/skills/develop develop && zip -r develop.zip develop # claude.ai → Customize → Skills → + → Create skill → Upload a skill → develop.zip
SKILL.md
View on GitHubDevelop
Own routing, integration, and user communication. Workers run one bounded stage or phase, then return without self-promotion.
Route
First apply change size. Record the size and honor request-specific step overrides before selecting stages.
researchonly for independent facts.innovateonly when alternatives add value.specwhen goal, requirements, acceptance, scope, approach, constraints, or phase boundaries are unsettled.implementfrom this run's decision-complete spec, or directly for narrow settled work.review-codeafter implementation validation only when the size policy or user requires it. Failed or unavailable validation does not waive required review.
Skip uninformative stages, but never the non-trivial-work spec sufficiency gate.
Continuity
This orchestrator owns transitions, including after spec's internal hard stop:
- After saving decision-complete
spec_<work-name>.md, summarize outcome, decisions, phases, touch points, and verification; immediately continue toimplementwithout waiting forapprove,implement, or/implement. - Pause only for an unresolved material product/architecture choice or required user selection. Ask once; never re-ask settled points.
- Workers remain in their assigned skill and return without self-transition. Select the next stage under the current unfinished request's authority. After completion, route each new request by its own size.
Worklog
For stages that need artifacts, create one .ai/worklog/<yyyyMMdd>_<work-name>/; pass its exact path and work-name throughout. Artifacts are optional except when their stage produces:
research_<work-name>.mdspec_<work-name>.mdtrash/
Dispatch
- Main owns decisions, communication, and routing. Researchers supply facts; bounded
specworkers may reason and author the artifact. - The
implementhost materializes each effective brief from the current shared contract and one phase delta, resolving requirements, criteria, verification, boundaries, mutation ownership, and selected tier. - Keep short phases on the host. When the size policy permits delegation, dispatch one implementor per ready phase. Run mutually declared
parallel_withphases together only when persisted phase records give disjoint mutation ownership, there is no dependency, and there is no shared mutable state. Do not infer safe parallelism from runtime plans. - Apply model selection to host reasoning and every worker. Preserve phase instructions and ownership; pass verified author provenance to research and spec workers.
- Assign one bounded stage/phase per worker, never the whole workflow; wait for all workers before integration.
- Include the phase outcome in that same dispatch brief. On Cursor, try once before the launch to create one native Goal for that outcome and say so in the brief; if the call cannot be made or is refused, the same brief tells the worker to create one Goal for its own agent. Codex and Claude put that outcome in the one parent checklist in the same brief and do not ask for a native Goal tool. Do not launch without the outcome text, do not respawn a worker because a Goal-gate phrase is missing, and do not interrupt a worker that has already started.
Validation
- Establish a baseline before implementation when practical.
- Verify phase criteria against the integrated current revision before completion; later writes invalidate affected evidence. Apply
implement's final verify gate, including whole-spec acceptance. - Run validation required by the size policy and repository. When review is required, select a model under
review-codeand dispatch a fresh independent reviewer with request, size, effective brief, resolved model/effort, repository instructions, diff, and validation evidence. The reviewer returns findings directly and creates no review file unless the user explicitly requested one. In Codex, use a fresh reviewer role, never a shared-agent plugin path; Claude Code/Cursor may usecode-reviewer. - Fix every
Openfinding throughimplement, using the host for short fixes. Rerun affected validation. Follow the size policy for required re-review, optional review consent, and at most two fix/re-review rounds. Each round must close at least one finding or add concrete evidence; otherwise stop as blocked. - Complete only when required validation and review pass, and every actionable finding is fixed or disproved. The host may confirm Low-only fixes for small work; label them host-verified. Prevention promotion requires the reviewer's fresh confirming review.
- Leave failed-phase continuation to the active
implementhost's failure-only bounded continuation policy. Apply the size policy to review of returned changes even while implementation remains blocked. - Report only the outcome, changed files, validation, actionable review findings, and blockers.
Files in this skill
3develop skill questions
How do I install the develop skill in Claude Code?
Run npx degit leszekgruchala/buddy/skills/develop ~/.claude/skills/develop. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the develop skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.