Project alignment verification for Claude Code. Checks whether your project is still heading where you intended, detects drift, diagnoses causes, and recommends corrections. Works standalone or with Discovery Orchestrator.
Course Corrector is the "check engine light" for projects. Run it at any point during a project to answer:
- Are we still solving the right problem?
- Has scope crept? In what direction?
- Are we building what we set out to build?
- What corrections are needed, and how urgent are they?
| Mode | When | Experts | Output |
|---|---|---|---|
| Quick | DO artifact < 30 days old, or user wants pulse check | 1 (alignment-auditor) | Health Check (1 page) |
| Full | DO artifact > 30 days old, no artifact, or user reports concerns | 4 (all) | Drift Report (comprehensive) |
| Phase | Name | What Happens |
|---|---|---|
| 0 | Check-in | Detect DO artifacts, select mode, confirm domain |
| 1 | Baseline | Establish the "north star" — original intent and success criteria |
| 2 | Reality | Survey current state — what's been built, decided, changed |
| 3 | Gap Analysis | Expert comparison of baseline vs reality |
| 4 | Drift Diagnosis | Why did drift occur? Categorize and root-cause (Full only) |
| 5 | Correction | Design go/pivot/stop corrections |
| 6 | Recalibration | Generate report, score quality, persist artifact |
| Expert | Role | Mode |
|---|---|---|
| alignment-auditor | Measures gap between intent and reality | Both |
| risk-sentinel | Evaluates risks of correcting AND not correcting | Full |
| drift-pathologist | Diagnoses why drift occurred | Full |
| trajectory-planner | Designs the correction path | Full |
Alignment Score (project health):
- Intent Fidelity (0.25) + Scope Adherence (0.20) + Constraint Integrity (0.20) + Success Trajectory (0.20) + JTBD Alignment (0.15)
Document Quality (report quality, same 5D as Discovery Orchestrator):
- Specificity (0.25) + Actionability (0.25) + Traceability (0.20) + Completeness (0.15) + Coherence (0.15)
Final Score = Alignment (60%) + Document Quality (40%). Pass threshold: 3.5/5.0.
Course Corrector works standalone but integrates natively with Discovery Orchestrator:
- Reads DO artifacts from
.discovery/as the "north star" baseline - Writes correction reports to
.corrections/with compatible format - Shares the same 6 domain adapters (software, product-ux, ai-projects, consulting, internal-tools, personal)
- Uses the same Quality Score footer format for document quality
Discovery Orchestrator → .discovery/artifact.md → Course Corrector → alignment check
↑ |
└───── re-discovery recommended if drift > threshold ─────┘
Course Corrector doesn't just wait for you to report concerns. Throughout every phase, it actively monitors for:
- Contradictions between current work and original intent
- Scope items never in the original plan
- Constraints being silently relaxed
- Emotional signals of uncertainty
When detected, it pauses and surfaces the finding: "I'm noticing something — can we check?"
Each adapter tailors drift detection and correction vocabulary to the domain:
| Domain | Key Drift Signals |
|---|---|
| software | Feature branching without merging, tech debt "won't fix", architecture decisions in PRs |
| product-ux | Features for wrong personas, success metrics redefined post-launch, competitor copycat |
| ai-projects | Eval dataset stale, prompt-as-architecture, cost-per-query over budget |
| consulting | Stakeholder alignment skipped, SOW scope expanding, deliverables drifting |
| internal-tools | Compliance deferred, adoption not tracked, exception handling still manual |
| personal | Decision deadline passed, analysis replacing action, question itself changed |
bash install.shInstalls to ~/.claude/skills/course-corrector/.
In Claude Code:
/course-corrector check my RAG project
/course-corrector [paste project context]
Or invoke when you have a Discovery Orchestrator artifact:
/course-corrector check
Course Corrector will auto-detect .discovery/ artifacts and use them as the baseline.
course-corrector/
├── SKILL.md # Orchestrator
├── phases/ # 7 phases (00-06)
├── experts/ # 4 expert agents
├── adapters/ # 6 domain adapters
├── templates/ # 3 output templates
├── scoring/ # Dual-track rubric
├── examples/ # 2 example sessions
├── install.sh # One-command install
├── README.md # This file
└── LICENSE # MIT
MIT