| name | claude-x |
| description | Claude-first development methodology. Pre-flight classification, pattern-based execution, outcome tracking. Use for any autonomous work requiring predictable quality. |
Claude-X
Methodology for predictable quality in autonomous work.
Core Loop:
- Pre-flight → Classify task, assess predictability
- Execute → Use known patterns, checkpoint constantly
- Post-mortem → Log outcome, update patterns
Session Start
Read this file (not from memory — compaction loses details)
Read project's session file (if exists)
- Location:
[PROJECT]/session/CLAUDE-DEV-PLAN.md - Read STATUS section — orient to current thread
- If no session file exists, create one from templates/session.md
- Location:
Add start pulse to session file (one line: time, "start", thread)
Check patterns at patterns/INDEX.md
If new autonomous work: Complete preflight.md classification
Pre-Flight (New Work)
Before ANY autonomous work:
- Open preflight.md
- Classify: pattern, scope, testability
- Assess predictability: HIGH / MEDIUM / LOW
- Check risk factors
- Decision: PROCEED / CAUTION / STOP
If LOW predictability or 4+ risk factors → STOP
- Split the task
- Clarify success criteria
- Escalate for guidance
Execution
Pattern Selection
Known pattern exists? → Use it
Novel but well-defined? → Attempt, document if succeeds
Unclear? → Clarify before starting
Checkpointing (Lightweight)
Primary rule: Every response that uses tools ends with a checkpoint line.
Format (inline in response, not in separate file):
---
CP: [phase] | [N] tools | [action completed] | next: [literal next step]
Example:
---
CP: Phase 4.2 | 6 tools | wrote App.tsx | next: run build to verify
Why this works:
- Zero file switching — checkpoint is in the response itself
- Forces articulation — stating "next step" catches drift before it happens
- Recovery-friendly — if context compacts, last CP shows exactly where stopped
Session File Updates
Update project's session file only on:
- Session start (add start pulse)
- Detected compaction/recovery
- Topic change (new task entirely)
- Session end (copy last few CPs as summary)
Do NOT update session file after every commit — that overhead causes skipped checkpoints under pressure.
Context Pressure Signals
- Response >1000 words
- Re-reading a file already read
- Uncertainty about what was done
- Tool calls >8 in response
When detected: End current response with CP line. Start fresh in next response.
Bail-Out Rules
- Same error twice → STOP, document blocker
- Three failed approaches → Escalate, no more guessing
- Scope creep detected → Checkpoint, reassess
Post-Mortem (Session End)
Complete postmortem.md — target <3 minutes.
Required:
- Log outcome to project's
outcomes/YYYY-MM.md - Copy last few CP lines to session file as summary
- Archive session file if significant work done
If variance:
- Note cause (one line)
- Create/update pattern if applicable
Project Setup
Projects adopting claude-x should create:
[project]/
├── session/
│ └── CLAUDE-DEV-PLAN.md # From templates/session.md
├── outcomes/
│ └── YYYY-MM.md # Monthly post-mortems
└── patterns/ # Optional: project-specific patterns
└── INDEX.md
Copy templates from this repo's templates/ directory.
Quick Reference
Predictability
| Level | Criteria | Action |
|---|---|---|
| HIGH | Known pattern + S scope + testable | Proceed |
| MEDIUM | Known pattern with risks OR novel but defined | Proceed with caution |
| LOW | Novel + large OR untestable | Stop, split, clarify |
Scope
| Size | Duration | Files | Action |
|---|---|---|---|
| S | <15m | 1-2 | Single session |
| M | 15-60m | 2-4 | Single session, frequent checkpoints |
| L | >60m | >4 | MUST split first |
Checkpoint Examples
CP: Phase 2.1 | 3 tools | created test file | next: run tests
CP: Phase 2.1 | 5 tools | tests passing | next: implement production code
CP: Phase 2.2 | 8 tools | hit error twice | next: STOP, document blocker
CP: recovery | 2 tools | re-read STATUS | next: continue Phase 3.1
Cross-Project Metrics
Claude-x performance is tracked per-project in each project's outcomes/ directory.
To evaluate methodology performance across projects:
- Check
adopters.mdfor list of projects using claude-x - For each project, fetch its
outcomes/directory - Aggregate SUCCESS/PARTIAL/FAILED rates
- Compare predicted vs actual accuracy
This query-time aggregation avoids cross-repo commits during execution.
Integration
With Claude Code
- Add symlink:
ln -s /path/to/claude-x ~/.claude/skills/claude-x - Use native async subagents (not custom scripts)
- Use
/statusfor context monitoring
With Project Dev Plans
Each project maintains its own CLAUDE-DEV-PLAN.md:
- STATUS section for recovery
- Micro-phases for execution
- Files list per phase