This plugin helps you write code like a seasoned engineer, even if you don't know all the best practices. It leans on our own thinking for reasoning, since even the best AIs cannot reason: they predict the next token in a sequence, and we have to decide which direction to point them in. It aims to slow the AI down when the scope is unclear and let it run fast when it knows what to do. At the same time, it catches the AI before it lets itself make mistakes or assumptions, so the pace always fits the level of understanding.
How it was built, and what it caught on the site it was tuned on, is told at dansabin.com/projects/ponderly.
Docs: guides to plan a year, build a feature, fix a bug, make a small change and tune a template; the Reference; the evidence; contributing.
ponderly is a plugin for Claude Code.
It adds skills, agents and team rules that every session loads. This
repository is its marketplace, sabind, and the marketplace holds one
plugin, ponderly. From any terminal:
claude plugin marketplace add sabind/ponderly
claude plugin install ponderly@sabindThe plugin loads the next time you start Claude Code. In a session that is
already open, run /reload-plugins.
You ask Claude for a job, and the plugin runs the skill that fits it. A small,
contained change goes to one agent that builds it and proves it. A bug goes to
bugfix, which reproduces it before any code changes. Anything bigger becomes
a spec of requirements, design and tasks, each of which you approve before the
next, and implement-spec builds it wave by wave. Before the pull request,
readable-code makes the branch's code read plainly without changing what it
does, and tighten-tests slims its tests without losing a covered line.
flowchart LR
strategy --> hld --> lld --> start-task
start-task -->|"one contained change"| task-executor
start-task -->|"a defect"| bugfix
start-task -->|"anything bigger"| spec["spec: requirements, design, tasks"]
spec --> implement-spec
implement-spec -->|"each task"| task-executor
implement-spec -->|"each wave"| wave-reviewer
start-task -->|"after a done change"| pr
implement-spec -->|"every task ticked"| pr
bugfix --> pr
readable-code -->|"then the tests"| tighten-tests
readable-code --> pr
tighten-tests -->|"code not yet readable"| readable-code
tighten-tests -->|"each gap found"| start-task
tighten-tests --> pr
review -->|"each lens"| review-arm
review -->|"the merged findings"| wave-reviewer
The Reference describes each skill and agent, including
report, which writes hand-backs like the one above.
- Claude Code without the plugin. Claude Code is an agentic coding tool that reads your codebase, edits files and runs commands, in the terminal, an IDE, a desktop app or the browser. Use the plugin on top of it when you want each request routed to a skill, specs you approve before code, and the same team rules in every session.
- Kiro. Kiro is an agentic IDE built around spec-driven development: it breaks complex features into detailed implementation plans with tracking, and its steering files hold a project's rules. In ten months of my own use it was good for thinking a feature through, but it often reported a fix or a task as done before checking it, and I had to restate rules in each workspace. The plugin brings the same spec way of working into Claude Code, ticks a task only after its checks have run, and loads the same rules in every repository.
make validate runs the free checks under checks/ and claude plugin validate, with no model call, and CI runs it on every pull request. The
make eval-* targets run claude plugin eval cases for the skills, agents,
hooks and most rules; each case is a real model call, so they run by hand and
never in CI, and the evidence shows what they found.
- It does not commit, push or open a pull request unless you ask, and
propens one only after you approve its draft.readable-codeandtighten-testsleave their changes uncommitted. - It does not commit on the default branch. When you are on it,
prandimplement-specoffer to create a branch first. - It does not decide what to build. You own each spec document and approve it
before the next, and
implement-specnever edits the spec beyond its checkboxes without your approval. - The plugin itself does not read a credential or send anything anywhere
except through
gitandghto your repository's own remote, and thereview-armagent's web searches, which name only a tool outside the repository and the behaviour in question, never your code. Thetask-executoragent and thereadable-codeandtighten-testsskills also run the project's own tests, and the two skills its other tools, which do whatever they do in that project.tighten-testsadds a coverage tool only when you say yes. What each part runs is in the Reference's What the plugin runs and sends.
To set up, test and submit a change, read the contributing guide.
MIT. See LICENSE.