Three files. One build.
Tripod is a minimal spec-driven workflow for Claude Code. No harness engineering, no hundred-markdown-file frameworks. A project stands on exactly three sources of truth:
| File | Role | Written by |
|---|---|---|
SPEC.md |
What to build — requirements, decisions, formulas, scope | /tripod:spec |
CLAUDE.md |
How to build it — stack, architecture, design language, code rules | /tripod:spec |
TASKS.md |
When — small, ordered, verifiable tasks with checkboxes | /tripod:tasks |
A tripod needs exactly three legs. More is clutter.
Agentic coding fails in two familiar ways: no plan (the agent improvises architecture per session and the codebase drifts) or too much plan (a framework of dozens of interlocking docs that burns the context window before work starts). Tripod sits at the minimum that works: the context stays concise, every session reads the same three files, and progress is a checkbox — visible in every diff.
In practice, this loop produces near-complete apps: the agent never relitigates decisions (they're in SPEC.md), never invents styles (they're in CLAUDE.md), and never wonders what's next (it's the first unchecked box in TASKS.md).
/plugin marketplace add ayberkyasa/tripod
/plugin install tripod
Describe your app in plain language. Claude interviews you in short rounds — architecture-defining decisions first, then the MVP boundary, then forks in the road — proposing defaults you can veto instead of making you design from scratch. It then writes SPEC.md and CLAUDE.md.
Then you read them. Human review of these two files is a mandatory step, not a formality. Edit anything; they're yours.
Claude reads both files and writes TASKS.md: phased, sequential tasks sized to one focused session each, every one with concrete file paths and a verifiable "Done when" condition. Foundation first, then the hard algorithms as pure tested logic, then services, then UI, then hardening.
Review the plan. Reorder, resize, veto.
Each invocation: find the first unchecked task → implement it under CLAUDE.md's rules → run its "Done when" literally → check the box → commit. Never a checked box on a red build. Repeat until the file is all [x].
- Every task has an objectively verifiable done-condition — a command that passes, a behavior you can observe. "Code is written" is not one.
- A task exceeding ~300 lines of new code gets split.
- Sequential by default;
[P]marks the rare parallel-safe task. - The three files are always truthful: when reality diverges (a constant changes, a constraint is discovered), the file is updated in the same session — with your confirmation.
- Deferred features are listed in TASKS.md, so no session "helpfully" implements them.
Isn't this just spec-kit / agent-os / …? Same family, different bet. Tripod bets that three files are enough and that anything beyond them costs more context than it earns.
Do I need the plugin, or can I just copy the templates?
The templates/ are plain markdown — copy them and run the process by hand if you prefer. The plugin adds the three slash commands and the discipline baked into them.
Does it only work for iOS/Swift? No. The skills are stack-agnostic; CLAUDE.md is generated for whatever stack you choose in the interview.
MIT © Ayberk Yasa