Skip to content

v0.1.0 — Initial release: claude -p /goal subprocess kernel

Choose a tag to compare

View all tags
@ulissesflores ulissesflores released this 17 May 04:55
· 21 commits to main since this release
v0.1.0
214cfa0

ASP v0.1.0 — Anticipating Shadow Points

First public release of the anticipating-shadow-points Claude Code skill.

What is ASP?

A Claude Code skill that orchestrates a 13-phase protocol for non-trivial engineering tasks. Invoked via /asp <task>. Combines Klein's pre-mortem (1998), Berkeley MAST 14-mode failure taxonomy, Plan-and-Act separation, and Anthropic's Plan-Validate-Execute discipline. Worker/evaluator execution via claude -p /goal subprocess.

Highlights

  • claude -p /goal subprocess execution kernel — empirically validated 2026-05-17 (Claude Code 2.1.143). Worker/evaluator separation via Anthropic's native primitive, invoked as child process from skill context.
  • Iron Law 11: NEVER trust `$?` from `claude -p` — agent returns exit 0 even when gracefully refusing the goal. Routing decisions MUST parse JSON output. (Empirically discovered via test C of the validation battery.)
  • 12 Iron Laws enforcing pre-mortem discipline, validator independence, contractual micro-TODOs, fresh acceptance-test evidence per step, per-deliverable sign-off, and memory write-back at project close.
  • File-based parent/child contract via `~/Developer/ASP/runtime//` — child writes `exec.json`, parent reconciles task statuses against filesystem evidence (Iron Law 4).
  • Native in-session kernel fallback for tasks <5 micro-steps or when `--no-claude-p` flag passed.
  • 5 structured evals + RED/GREEN/REFACTOR baseline tests — empirical evidence that the skill raises shadow-point coverage from ~47% (baseline agent) to ~100% (with skill loaded).
  • Multilingual docs: EN (native) + ES/PT/IT/HE (machine-assisted, PRs for native review welcome).
  • Idempotent install/uninstall scripts + verify.sh with 20 success criteria.

Quick start

```bash
git clone https://github.com/ulissesflores/anticipating-shadow-points.git ~/Developer/ASP
cd ~/Developer/ASP
./scripts/verify.sh --pre-install # 19/20 OK (skips staging check)
./scripts/install.sh --dry-run # preview
./scripts/install.sh # install to ~/.claude/skills/
```

Then in any Claude Code session: `/asp `.

Empirical evidence shipped

  • `tests/claude-p-goal-runner-probe.md` — battery A/B/C/D documented verbatim
  • `tests/goal-invocability-probe.md` — in-session limitation + subprocess discovery
  • `tests/baseline-pressure-tests.md` — RED phase: 3 scenarios run without the skill, violations documented
  • `tests/evals-summary.md` — 5/5 evals at 100% coverage (≥80% threshold cleanly exceeded)
  • `tests/criteria-final-checklist.md` — all 16+ verify criteria documented with evidence paths
  • `tests/advisor-final.md` — 3 advisor calls log + final verdicts

Dependencies

  • Required: Claude Code 2.1.139+ (for `/goal` slash command in `claude -p` mode)
  • Required: `bash`, `jq`
  • Optional: `~/.agent/` agentic-stack (enables `recall.py` lessons retrieval; skipped silently if absent)

License

MIT.

Acknowledgements

This skill exists thanks to an iterative dialog between user and agent that included three architecture pivots (v3 → v4 → v5), empirical testing initiated by the user (battery 2026-05-17), and advisor-caught bugs pre-ship. The most important architectural finding (`claude -p /goal` subprocess as primary kernel) came from the user's instinct to test rather than accept the initial "we can't invoke /goal" conclusion.

Sources studied:

  • Klein, G. (1998, 2007). "Performing a Project Premortem". Harvard Business Review.
  • Cemri et al. (2025). "Why Do Multi-Agent LLM Systems Fail?" — arxiv 2503.13657.
  • Erdogan et al. (2025). "Plan-and-Act: Improving Planning of Agents". arxiv 2503.09572.
  • Shinn et al. (2023). "Reflexion".
  • Anthropic Claude Code 2.1.139 /goal command (released 2026-05-12).
Assets 2
Loading