Ship ambitious work without missing the obvious.

ASP is a Claude Code skill that surfaces the failure modes a senior engineer would catch in a pre-mortem — before any code is written — through a forced MAST-14 pre-mortem, a prompt-isolated independent validator, and a claude -p /goal subprocess executor.

Install in 30 seconds → View on GitHub See the evidence ↓

Evidence — preliminary

preliminary

5 tasks × 2 conditions, one dispatch each,
single model, author-adjudicated
(adjudicator readings in benchmark/, §5)

runtime invariants ("Iron Laws")
enforced by the skill

structured evals
across real domains

The evaluation is deliberately small: 5 tasks, two conditions, one dispatch each, a single model, author-adjudicated. A deterministic adjudicator shipped in benchmark/ — which anyone can re-run — re-scores the same transcripts under an author-defined criterion; its readings are reported descriptively in the whitepaper (§5), not as a coverage estimate against external ground truth. The sample size and non-blind adjudication are stated limitations (§7).

Same task, two outcomes

Real subagent dispatches from the 2026-05-17 RED/GREEN baseline. Task: “Add a NOT NULL column tier (default ‘free’) to user_profiles. Table has 1M rows in production.”

❌ Without ASP — baseline agent

ALTER TABLE user_profiles
  ADD COLUMN tier text
  NOT NULL DEFAULT 'free';

CREATE INDEX CONCURRENTLY
  idx_user_profiles_tier
  ON user_profiles (tier);

~60% shadow-point coverage. Misses: replica lag, app deploy ordering, monitoring spike, trigger interaction, foreign key cascade.

✅ With ASP — same agent, same task

Pre-mortem (30 days out):
 1. Full table rewrite locks
 2. RLS policy not updated
 3. Stale TS client types
 4. Service-role bypass risk
 5. Replica lag during DDL
 6. App deploy ordering
 7. Monitoring spike
 8. Backup window collision
 9. Index INVALID state
10. WAL/disk pressure
11. CDC realtime flood
12. Rollback impossible
13. CPU credit exhaustion
14. Audit trail missing
15. Time-zone in cron logic

Illustrative: all 15 surfaced on this example task. Each surfaces a mitigation before any SQL runs.

Built on established methodology · Klein pre-mortem · Berkeley MAST 14-mode · Plan-and-Act · Reflexion · preliminary evaluation (5 tasks), 2026-05-17 (see whitepaper §7 for limitations)

What ASP does

🎯 Pre-mortem before code

"Imagine the failure. Now describe why." Prospective hindsight — Mitchell, Russo & Pennington (1989) showed it generates ~30% more reasons than imagining the same event as merely possible; Klein (2007, HBR) operationalized it as the pre-mortem. ASP runs it on every non-trivial task.

🔍 MAST 14-mode checklist

Berkeley's failure taxonomy forces coverage of less-obvious categories — time/timezone, observability, op-ex, contract drift — that free-form pre-mortem clusters away from.

🛡️ Independent validator

A fresh-prompt subagent audits the plan before execution. Self-critique fails on high-confidence hallucinations; prompt isolation breaks that collusion. Iron Law 2.

⚡ `claude -p /goal` kernel

Spawns Anthropic's official worker/evaluator kernel as a subprocess. Empirically validated 2026-05-17. Iron Law 11: never trust $? — parse JSON.

See it run

$ /asp Add a NOT NULL column `tier` to user_profiles. 1M rows.

Phase 1 — Parallel research (codebase + web + recall.py)          ✓
Phase 2 — Shadow-point detection (Klein + MAST 14-mode)            ✓ 14 found
Phase 4 — Macro plan + Deliverables Register (D01..D08)            ✓
Phase 5 — Independent validator (separate-prompt subagent)         APPROVE
Phase 7 — Awaiting user approval of charter + plan + deliverables  ✓
Phase 8 — Micro-TODO contract emitted: 23 tasks (PRE/ACT/POST/ACC/FAL)
Phase 9 — claude -p /goal autonomous execution                     → running

Goal Spec: All 23 tasks completed with ACCEPTANCE-TEST evidence;
           all 8 deliverables marked `aceito`; 12 Iron Laws respected.
Hard stops: turn > 80 · elapsed > 30min · 3 consecutive falsifications.

Every step is auditable. Every task has a verifiable acceptance test. Nothing ships without fresh evidence.

Get started

claude plugin marketplace add ulissesflores/anticipating-shadow-points
claude plugin install anticipating-shadow-points@anticipating-shadow-points

Invoke /anticipating-shadow-points:asp <task> in any Claude Code session. Two more install paths (dev mode --plugin-dir, standalone install.sh for bare /asp) are in the GitHub README.

Requirements: Claude Code 2.1.139+, bash, jq.

When to reach for ASP

✅ Use ASP for	❌ Skip ASP for
Features touching data or shared state Migrations (schema, dependency, API) Refactors across many files Deploys with external dependencies Architecture decisions Root-cause debug investigations	Typo fixes Single-line edits Read-only questions ("what does X do?") Tasks you can finish in under 5 minutes

Ready?

claude plugin marketplace add ulissesflores/anticipating-shadow-points
claude plugin install anticipating-shadow-points@anticipating-shadow-points

Install now Star on GitHub ⭐

Maintainer: Carlos Ulisses Flores — CTO & Chief Researcher at Codex Hash Research Laboratory · MSc AI candidate · São Paulo, Brazil · ORCID 0000-0002-6034-7765 · Lattes 6905246706890561 · @ulissesflores

Languages: 🇬🇧 English · 🇪🇸 Español · 🇧🇷 Português · 🇮🇹 Italiano · 🇮🇱 עברית

Deep dive: ARCHITECTURE.md — 10-part SOTA design document (academic background, architectural journey, empirical discoveries, derivation of all 12 Iron Laws).

Cite this work: CITATION.cff · Community: Contributing · Code of Conduct · Security · Changelog

License: MIT