ideasift

name	ideasift
description	IdeaSift — Survivor Idea Filtration Tool. Named AI personas adversarially filter ideas against a corpus. Multi-roster comparison for cross-panel validation.
triggers	["ideasift","ideasift ideas","filter ideas","adversarial review","survivor document","review architecture","challenge ideas","sift this","gap analysis"]

IdeaSift — Survivor Idea Filtration Tool

🛠️ FIRST-RUN — check install BEFORE invoking anything

This skill calls an ideasift binary that ships in COR-CODE source-only. Always run this check first — don't assume the binary exists, don't invoke it blind.

which ideasift

If it returns a path → tool is installed, proceed.

If it returns nothing → tool is NOT installed. Walk the user through install BEFORE attempting any filtration:

"ideasift isn't installed yet. The source ships in COR-CODE at tools/ideasift/. Install steps:

1. cd <COR-CODE-root>/tools/ideasift
2. npm install (fetches deps incl. Anthropic SDK)
3. npm run build (compiles TypeScript → dist/)
4. npm link (symlinks ideasift to your global npm bin)
5. Verify: which ideasift && ideasift --version

Want me to run those steps for you, or would you rather run them yourself?"

Don't run the install silently — npm install fetches third-party packages and npm link modifies a global PATH-resolved binary. The user should know.

Pre-requisite — env: ANTHROPIC_API_KEY must be set (personas are Claude-driven). Check with [ -n "$ANTHROPIC_API_KEY" ] && echo "set" || echo "missing" before invoking. Warn the user if missing — do not pretend the run will work.

---

Named AI personas generate ideas, score them asymmetrically, and adversarially challenge each other. Only survivors make the final document. Multi-roster comparison mode cross-references what survives across different team compositions.

bughunt finds bugs in code. ideasift finds weak ideas in architecture.

Karpathy Principles Applied

These principles (from autoresearch analysis) improve how IdeaSift is used:

The Complexity Tax

When evaluating survivor ideas, apply this rule before accepting:

• An idea that improves the architecture but adds significant complexity? Questionable.
• An idea that improves the architecture by REMOVING something? Always keep.
• An idea that improves nothing but simplifies? Keep.

Ask for every survivor: "Is the improvement worth the complexity it adds?" Renzo's territory but ALL personas should weigh this.

Eval-Locked Mode

When running IdeaSift iteratively (sift → implement survivors → sift again), the SCORING FUNCTION must not change between runs. The 8 stages and +2/-3 asymmetric scale are the immutable eval. If you change scoring between iterations, you can't compare results. Lock the eval, change the corpus.

Context Hygiene

IdeaSift output can be verbose. When the project Claude reads the survivor document:

• Read the carry_forward items only (not the full report)
• Ignore discarded items entirely (they failed for a reason)
• For worth_discussion items, read the challenge arguments before deciding

Don't dump the entire survivor-document.json into context. Extract the signal.

Recovery When No Ideas Survive

If IdeaSift returns 0 survivors, don't conclude "the corpus is perfect." Instead:

1. Run again with different roster (--compare mode if not already)
2. Try with duplicate personas (Renzo,Renzo forces within-territory competition)
3. Narrow the corpus to a specific section and sift that
4. Ask: is the corpus so vague that nothing is falsifiable? That's the real problem.

The Loop Pattern

For iterative improvement of a document:

1. Sift the document → get survivors
2. Apply survivors to the document
3. Sift again with same eval (same scoring, same stages)
4. If new survivors emerge → repeat
5. If 0 survivors → document has converged

Each iteration should produce FEWER survivors than the last. If iteration 3 produces more than iteration 1, something is wrong — you're introducing complexity faster than you're resolving gaps.

Location

Tool installed at: ~/.claude/tools/ideasift/

How to Run

Review a single document

ideasift ./architecture.md

Review a directory of docs

ideasift ./02-COMPLETED/ --budget 8

Compare mode — 3 roster compositions, cross-reference survivors

ideasift ./architecture.md --compare --budget 15

Custom roster — double up personas for within-territory competition

ideasift ./design.md --personas "Renzo,Renzo,Kael,Kael,Soren"

Quick mode (skip adversarial challenge stage)

ideasift ./PHASES.md --quick --verbose

Named Personas

7 constructed identities with territory allocation (reduces correlated output):

Name	Territory	Background
Mara	Ownership, SLA, governance	VP Ops Big Four → COO fintech
Kael	Security, legitimacy, provenance	Cloudflare security → UK financial services
Soren	State machines, replay, spec quality	Stripe payments infra → distributed systems
Lena	Trust, adoption, behavioural signals	Intercom product → B2B AI adoption
Renzo	Dead weight, overbuild, deletion	Two failed CTOs → independent advisor
Sage	Continuity, ambient awareness	Cognitive science → ambient computing research
Voss	Competitive attack, proof, speed	YC startup → platform engineering

Duplicate personas compete within territory — Renzo,Renzo creates Renzo-1 and Renzo-2, each forced to find what the other missed.

Roster Presets (for --compare)

Preset	Composition	Bias
hardening	Mara, Kael, Kael, Soren, Renzo	Safety, rigour, subtraction
product	Lena, Lena, Sage, Voss, Renzo	Trust, adoption, competition
full	All 7	Balanced coverage

Key Flags

Flag	Purpose	Default
--personas <names>	Comma-separated persona names (duplicates allowed)	all 7
--compare	Run all 3 presets and cross-reference	off
--rosters <presets>	Specific presets for comparison	hardening,product,full
--max-ideas <n>	Max ideas per persona	5
--challengers <n>	Challengers per idea	2
--budget <usd>	Max API spend in USD	5.00
-t, --thesis <text>	Non-negotiable thesis — personas improve it, never replace it	none
-q, --quick	Skip adversarial challenge	off
-v, --verbose	Show persona reasoning	off
-o, --output <dir>	Report output directory	./sift-report
-p, --provider <name>	anthropic or openai	anthropic
-m, --model <model>	Generator model	claude-sonnet-4-6
-c, --challenger-model	Model for challenges	same as generator
-d, --description <text>	Review description for report	Architecture review

API Keys

Set ANTHROPIC_API_KEY or OPENAI_API_KEY in environment, or pass --api-key.

Output

Two report files in the output directory:

• survivor-document.json — structured data (ideas, scores, challenges, stats)
• survivor-document.md — readable markdown with carry-forward, discussion, and discarded sections

Compare mode additionally produces:

• <preset>/ subdirectories with per-roster reports
• comparison.json — cross-roster agreement stats

Scoring System

Score	Meaning
+2	Genuinely new, high-leverage addition
+1	Valid sharpening of a partially-present concept
0	Interesting but low-leverage
-1	Vague philosophical restatement
-2	Duplicate of existing corpus presented as new
-3	Misleading or already-covered advice sold as a major gap

Duplicate personas in same territory: repeating another's idea costs -2.

The Method

1. Blind Generation: Each persona independently generates ideas within their territory
2. Novelty Screen: Check against corpus — already_covered / partially_present / real_gap
3. Asymmetric Scoring: Ideas scored +2 to -3 based on value and novelty
4. Adversarial Challenge: Other personas try to kill each idea using 6 gates (novelty, consequence, failure relevance, evaluability, layer confusion, counterfactual)
5. Survivor Synthesis: Only ideas that survive all stages make the final document

Typical Usage from Claude

Review an architecture document

User: "sift this architecture"
Claude: ideasift ./02-COMPLETED/DD-AUTONOMOUS-ORCHESTRATOR-architecture.md --budget 5 -o /tmp/sift-report
Claude: Reads survivor document, presents carry-forward items, discusses with user

Compare mode on a build plan

User: "run sift comparison on the phases"
Claude: ideasift ./PHASES.md --compare --budget 12 -d "COR Intelligence build plan review"
Claude: Reads comparison, highlights universal findings (survived all 3 rosters)

Quick gap check before building

User: "any gaps in this design?"
Claude: ideasift ./design-doc.md --quick --personas "Kael,Soren,Renzo" -o /tmp/sift-report
Claude: Reads survivors, flags genuine gaps, ignores discarded noise

Cost Expectations

Configuration	Typical Cost	Duration
3 personas, quick mode	$0.50-1.50	2-5 min
7 personas, full challenge	$3-6	10-20 min
Compare mode (3 rosters)	$10-18	30-60 min
Custom 5 personas with duplicates	$2-4	5-15 min

Rebuilding

If the tool needs updating:

cd ~/.claude/tools/sift
# Make changes to src/
npm run build