What is Solo SDLC
Ship reliable software, alone.
An AI-native development pipeline for solo builders. Structured phases. Human gates. Agent-legible artifacts. Designed for human builders — and legible enough that an AI agent can drive it.
- Phases
- 7
- Specialists
- 16
- Lifecycle hooks
- 6
- Self-test assertions
- 50
Why it exists
Solo developers ship alone. No code reviewers, no QA team, no DevOps engineer. Every decision falls on one person.
AI coding without structure decays. Code that works today, breaks tomorrow. No consistency, no process, no learning between features.
Enterprise SDLCs crush solo builders. Built for teams of fifty. You need the discipline — without the bureaucracy.
What's in the toolkit
Installed by sdlc.sh install. Three profiles: minimal, core (default), full.
Pipeline
- 7 phases — Discover → Architect → Plan → Build → Verify → Ship → Retro, with a 1.5 detour for UI
- Human gates — Every phase has an explicit approval step you control
- Per-feature folders — Every phase artifact lives in docs/sdlc/yyyy-mm-dd-name/
Workers
- 16 specialists — Test, security, perf, database, root-cause, linter, auditor, …
- 2 review subagents — typescript-reviewer + silent-failure-hunter; read-only second opinions
- 9 slash commands — /new-feature, /status, /gate-check, /checkpoint, /pr, …
Guardrails
- 6 lifecycle hooks — SessionStart, pre-commit lint+tsc, console.log warner, Stop pattern extractor, …
- Compliance audits — 23-rule compliance check before every commit
- Codebase auditor — OWASP Top 10, CWE, SANS/CIS, language idioms
Memory
- State snapshots — .claude/snapshots/ — pre-compact state, checkpoints, pattern candidates
- Bug log + backlog — Read before every build prompt; never re-introduce a fixed bug
- Retro pattern log — Recurring issues flagged at 3+ occurrences for systemic fixes
Integrations
- 5 MCP servers — Vercel, GitHub, Context7, Firecrawl, Chrome DevTools — live data, not training memory
- Vercel-native deploy — Zero-touch deploy on merge; runtime logs read from the deploy verification step
- GitHub-native flow — Issues, PRs, comments, CI all driven from inside the pipeline
Quality
- 50-assertion self-test — scripts/test/run-tests.sh runs before every toolkit release
- Versioned releases — v8.0 ships with release notes, version.json, and a single tarball
- Profile-based install — Pick your scope — minimal (20 files), core (39), or full (52)
How it works
Four layers, one direction. Each layer reports up and writes down.
Approve briefs, TDDs, plans. Merge PRs. Confirm destructive migrations.
Seven phases with explicit gates. One agent per phase, one job each.
16 specialists, 2 review subagents, 9 commands, 6 hooks. Each invoked by a phase.
Every action writes to .claude/snapshots/ + docs/sdlc/. Verification reads files.
Enriched throughout by MCP servers: Vercel, GitHub, Context7, Firecrawl, Chrome DevTools — live data, not training memory.
Built for two readers
The same artifacts serve both. A solo builder reads them as documentation. An AI agent reads them as instructions.
- You make the calls. Agents propose; you approve. Nothing ships without your sign-off.
- No process overhead. The pipeline is opinionated. You don't pick the steps; you walk them.
- Discipline by default. Every feature gets the same TDD, the same QA, the same release note. Repeatable beats brilliant.
- Your pipeline learns. Retros write back into CLAUDE.md and SKILL.md. The third strike becomes a checklist gate.
- Skills are instructions. Every phase ships as a SKILL.md with explicit steps, narration protocol, and gate checklist.
- State is in files. No hidden context.
sdlc-state.json,features.json, snapshots — all readable, all writable, all diffable. - Gates are deterministic. Lint, tsc, tests, compliance, security — every gate is a command that exits 0 or non-zero.
- The toolkit tests itself. 50-assertion self-test suite covers every hook, every profile, every error path. Regressions get caught before users.
What we believe
Structure beats talent
A disciplined process produces better results than raw skill. Repeatable beats brilliant.
Humans decide, agents execute
AI handles the grunt work. You handle the judgment calls. Nothing ships without your approval.
Write it down
Every phase produces artifacts. Context is documented, not remembered. Snapshots over screenshots.
The toolkit audits itself
50-assertion self-test suite. Every hook, every install profile, every error path runs before a release.
Get Started
Download the toolkit. Run the installer. Your next Claude Code session gains structure.