Based on my experience so far, here's what I would do on the heels of a fresh PAI v5.0 to fix some of the issues I've encountered with the release. Sharing personal experience in case it is helpful to others – no claims to being "authoritative", so apply this at your own risk. Most of what's written below is by my DA, reviewed of course. This document may evolve in time.
Go through these one by one with your DA. We've tried to put them in such an order that application of one fix will not run into issues or gaps addressed by later fixes in the list, but it's not 100% certain. Worth double-checking and understanding what you're doing (ask your DA to explain). Issues, Discussions and PR references usually give enough information for the DA to understand what needs to be done, explain what's going on, and make the fixes. It's possible to do them manually too, and for some simple ones it might be the easiest path. Do not apply blindly! Do the fixes before migrating data or playing around with the system too much – less frustration, and less misplaced odds and ends to clean up!
The contents of the DOCUMENTATION folder will come in handy.
How to apply patches: ~/.claude/ is not a git repo on a default install (through installer), so git apply / gh pr checkout don't work. Application is manual — Read the PR diff, Edit/Write the affected files.
1. Disable harness auto-memory. · ref: #1161 (two halves — harness-side here in #1, PAI-side in #2)
Claude Code's harness auto-writes rules into ~/.claude/projects/<dir>/memory/, which bypasses PAI's force-loading and structural enforcement. The v5 system prompt designates that surface the wrong home for rules.
→ Add "autoMemoryEnabled": false at the top level of ~/.claude/settings.json.
Now in place: when the DA captures feedback for improvement, it doesn't get lost in a black hole never to be looked at again.
2. Stop the Migrate skill writing rules to harness memory. · ref: #1161
The /migrate skill routes "AI collaboration rules" straight into harness memory on its own — re-creating the violation #1 was meant to prevent.
→ No upstream PR yet. Full recipe in REPAIR-LOG.md entry 2026-05-13.
3. Fix Migrate-approve TELOS path.
MigrateApprove.ts writes TELOS files to bare TELOS/ instead of USER/TELOS/. With this and #2 unfixed, /migrate will misroute both rule-shaped and TELOS-shaped content the first time it runs.
→ Apply upstream PR #1177.
Now in place:
/migrateis safe to run — rules go to PAI's canonical homes (CLAUDE.md, hooks,settings.json, or a skill's Gotchas section), and TELOS-classified content lands underUSER/TELOS/. Hold off on actually running/migratefor now, though — a few more things to do first.
4. Strip the full CLAUDE_CODE_* env-var family in Inference.ts. · ref: #931, PR #934 (closed unmerged; valid patch)
Latent on Claude Code <2.1.71 — surfaces the moment the CLI is updated. Problem manifests as PromptProcessing classifier timeouts at the 25s ceiling and fail-safe MODE: ALGORITHM | TIER: E3 fallbacks on every prompt.
→ Apply PR #934. 4-line patch to ~/.claude/PAI/TOOLS/Inference.ts. PR was closed on an apparent diff misread; the actual point is stripping the new CLAUDE_CODE_* family.
Also change timeout from 25s to 45s in PromptProcessing.hook.ts (see discussion #1112) because it's a bit tight for Sonnet.
Now in place: no more timeouts and classifier (which "mode" PAI will respond in) warnings. (Still needs to be confirmed that it really works.)
5. Wire 11 dormant hooks into settings.json (incl. SmartApprover).
v5.0 ships 11 hook files (PostToolUseFailure, Elicitation, InstructionsLoaded, PostCompact, TaskCreated, StopFailure, ConfigChange, FileChanged, TeammateIdle, PermissionRequest, CheckpointPerISC) that aren't registered in settings.json. They exist on disk but never fire. Critically, this includes the PermissionRequest hook that runs SmartApprover.hook.ts — without it wired, every routine internal write (Algorithm ISA scaffolds, MEMORY/WORK/ edits) triggers a manual permission prompt.
→ Apply upstream PR #1200. Adds 10 new event blocks and wires AgentInvocation.hook.ts + CheckpointPerISC.hook.ts.
Now in place: the hooks v5.0 actually ships with are doing their job — per-ISC checkpointing, deferred-restore, doc-integrity checks. And SmartApprover handles routine permission prompts so the remaining fixes don't drown you in approval dialogs.
6. Trim skill descriptions under the global token budget. · ref: #1205
v5.0's skill descriptions are too long. Their combined length exceeds Claude Code's total budget, so descriptions past the cap don't load — /doctor reports them as dropped, and the affected skills become invisible to the routing layer.
This is the chunkiest fix in the list — requires editing the description: frontmatter on every individual SKILL.md under ~/.claude/skills/.
→ Trim each skill's description: frontmatter. Preserve USE WHEN keywords.
→ See the Annex at the end of this file for proposed shorter descriptions, copy-pasteable into each skills/<Name>/SKILL.md frontmatter. Total is ≈26,000 chars (down from ≈59k);
7. Apply ISA Scaffold path fix.
ISA Scaffold workflow uses inconsistent paths — sometimes relative, sometimes absolute. The DA scaffolds an ISA at the start of every non-trivial Algorithm run, so without the fix those scaffolds can land in the wrong directory or fail outright.
→ Apply upstream PR #1214 — fixes two lines in ~/.claude/skills/ISA/Workflows/Scaffold.md.
Now in place: every installed skill is visible to the routing layer, and the ISA skill (which the Algorithm leans on at the start of every non-trivial run) scaffolds to the right location.
8. Fix Interview skill stale TelosRenderer reference.
Interview SKILL.md references TelosRenderer.ts in two places. The tool was removed before v5.0 shipped; TelosSummarySync hook does its job now. Step 5 of the workflow fails silently.
→ No upstream PR yet (tracking issue #1187). Drop Step 5 of the Interview workflow; rewrite the description to point at TelosSummarySync.hook.ts.
9. Apply InterviewScan bootstrap-template detection fix.
InterviewScan should flag unmodified bootstrap templates as incomplete so /interview prompts you to fill them. Without this, the scanner treats default-template files as if they were already authored, and Interview never prompts for the fields the user actually needs to set.
→ Apply upstream PR #1174.
Now in place:
/interviewruns cleanly. At this stage you can get started with/interviewin a separate session — go through a couple of documents in each sitting. The items below restore those files across compactions and sessions once Interview has produced them.
10. Audit Knowledge files for required frontmatter. · ref: no public source found — candidate for a fresh upstream issue
v5.0's LifeOsSchema.md v1.0 requires type, description, created, updated, quality, related on every MEMORY/KNOWLEDGE/ file. KnowledgeGraph and Knowledge harvester silently skip files missing these fields. Doesn't block /migrate or /interview (they don't read existing Knowledge files), but the graph silently under-reaches until you fix it.
→ No upstream PR. Audit every file under ~/.claude/PAI/MEMORY/KNOWLEDGE/; add minimal frontmatter to any file missing it. In our install, 14 files needed it.
Now in place: the Knowledge graph reaches every file.
11. Resolve DA identity via _registry.yaml.
If you've named your DA via /interview, the DA identity might live at USER/DA/_registry.yaml rather than legacy flat USER/DA_IDENTITY.md. v5.0's restore logic only reads the legacy path. Fix this before #12 so the restore reads from the right place.
→ Apply upstream PR #1216 — adds registry resolution with legacy fallback.
12. Restore identity context after compaction.
v5.0's RestoreContext.hook.ts doesn't include USER/PRINCIPAL_IDENTITY.md or USER/DA_IDENTITY.md in the post-compact restore. After a compaction event, the model loses identity context.
→ Apply upstream PR #1212.
13. Persist session handoff to disk on compaction. · ref: discussion #1099
v5.0's PreCompact hook outputs handoff context only to stdout. The advisory is visible inside the post-compaction window, but if the session ends afterwards, that context is gone for next time. Smoothing fix rather than a real-brokenness repair — without it, you can still resume work by reading the relevant ISA or asking the DA "what were we doing yesterday." With it, the recent handoff surfaces automatically at session start.
→ No upstream PR yet. Modify the PreCompact handler to write MEMORY/STATE/HANDOFF.md to disk alongside the stdout advisory. The SessionStart pipeline reads it on next launch.
Now in place: compaction events stop eating context — identity files (yours + the DA's) survive across the compaction boundary, and the handoff note persists to disk for the next session. End of the core flow — what's below is conditional (Pulse if you use it, cosmetic, voice).
Order matters here too: each fix unblocks the next. Get Pulse running first, then the dashboard, then the polish.
14. Add missing Pulse daemon dependencies.
Pulse won't start without yaml, minisearch, smol-toml packages.
→ Apply upstream PR #1125.
15. Fix Pulse path casing for Linux. [Linux only]
Dashboard 404s on Linux due to Pulse vs PULSE path casing inconsistencies. Invisible on macOS APFS — skip on macOS.
→ Apply upstream PR #1175.
16. Wire Pulse TELOS dashboard to real data. Pulse TELOS dashboard out of the box shows sample data rather than your TELOS. → Apply upstream PR #1147.
17. Apply Pulse TELOS dashboard bug fixes. Three known dashboard bugs (only relevant once #16 makes the dashboard show real data). → Apply upstream PR #1165 — three fixes + 113 tests + contract docs.
18. Add crash guards for one-mission users. Pulse Life Dashboard crashes on installs that have only one mission defined in TELOS. Only matters once the dashboard is rendering your data. → Apply upstream PR #1154.
19. Auto-start Pulse user-index module. Pulse user-index module doesn't start automatically with Pulse. → Apply upstream PR #1197.
20. Fix RepeatDetection state timing.
The DA can fall into tool-call loops — same tool with same arguments fired repeatedly within one response, spinning instead of escalating. Vanilla v5.0's repeat-loop detector saves state before the response, so it can't catch same-turn re-fires. With the fix, the detector kicks in mid-response and stops the loop. Polish rather than a real-brokenness repair — a new user won't inevitably hit this; it surfaces in longer multi-step Algorithm runs that get stuck retrying the same broken approach.
→ Apply upstream PR #1156. Three-file patch: rewrite the existing hook, add RepeatDetectionPromote.hook.ts, update settings.json registration.
21. Remove deprecated agents from the permission allow-list.
v5.0's settings.json allows Agent(QATester), Agent(BrowserAgent), Agent(UIReviewer), Agent(Pentester). All four are marked DEPRECATED in the system prompt. Nothing fails if left in (the agents still execute if invoked); the system prompt just says don't invoke them. Pure cleanup.
→ Apply upstream PR #1211.
22. Fix Daniel Miessler blog URLs in PAISystemArchitecture.md.
v5.0 ships example.com placeholder URLs in PAISystemArchitecture.md and CLAUDE.md.
→ Apply upstream PR #1188.
We didn't apply these — we don't use voice features here. They look reasonable on the PR pages and would matter for anyone running PAI with TTS / browser audio.
- PR #1101 — VoiceServer: local TTS support, voice personalities, automatic ElevenLabs fallback.
- PR #1116 — Pulse: WebSocket browser audio streaming (note: PR is Linux-only; macOS
afplaypath explicitly unchanged).
- 2026-05-14 — v1.0 initial publication
Copy each line into the corresponding skills/<SkillName>/SKILL.md frontmatter description: field.
Agents: Compose CUSTOM agents from Base Traits + Voice + Specialization, and manage 8 predefined functional TEAMS (engineering, security, design, etc.). ComposeAgent.ts merges base + user config, outputs unique prompt + ElevenLabs voice + prosody. Observer team variant: read-only oversight agents voting continue/halt/escalate against the tool-activity audit log. USE WHEN create custom agents, spin up agents, compose agent, spawn parallel agents, engineering team, security team, get the team on this, observer team. NOT FOR ad-hoc swarms or TeamCreate coordination (Delegation), single-threaded delegation without unique identities (Delegation Task).
ApertureOscillation: 3-pass scope oscillation: hold question constant, shift zoom — narrow tactical → wide strategic → synthesis. Delta between views surfaces design tensions invisible at any single scope. Requires two distinct inputs. USE WHEN aperture oscillation, zoom in and out, tactical vs strategic, architecture decision, feature fits system, local vs global design, wrong scope, scope negotiation. NOT FOR lens rotation (IterativeDepth), idea generation (BeCreative), causal chains (RootCauseAnalysis), or assumption decomposition (FirstPrinciples).
Aphorisms: Curated aphorism collection with full CRUD: content-based matching, themed search, thinker research, usage-history tracking to prevent repetition. Four workflows: FindAphorism (analyze content, return 3-5 ranked recommendations), AddAphorism (parse + theme-tag + dedupe), ResearchThinker (deep research, add sourced quotes), SearchAphorisms (by theme/keyword/author). USE WHEN aphorism, quote, find a quote, research thinker, add aphorism, search aphorisms, quote for newsletter, what did X say about, quote collection. NOT FOR general creative writing or social media post generation.
Apify: Scrape social media (Instagram, LinkedIn, TikTok, YouTube, Facebook), business data (Google Maps), e-commerce (Amazon), and general multi-page web (custom pageFunction) via Apify actors. File-based TypeScript wrappers filter and transform data **in code before returning to model context** — 95-99% token savings vs direct MCP. Parallel multi-platform queries via Promise.all for social listening + lead-enrichment dashboards. USE WHEN scrape Instagram, scrape LinkedIn, scrape TikTok, scrape YouTube, scrape Facebook, Google Maps leads, Amazon reviews, social listening, lead generation, web crawl. NOT FOR X/Twitter bookmarks, progressive scraping (BrightData).
Art: Generates static visual content across 20+ formats via Flux, Nano Banana Pro (Gemini 3 Pro), GPT-Image-1. 20 named workflows including Essay (blog headers), Mermaid, TechnicalDiagrams, D3Dashboards, Frameworks, Comics, YouTube thumbnails, PAI pack icons, RemoveBackground, Visualize (catch-all). Output staged to ~/Downloads/ for preview. USE WHEN art, illustration, diagram, flowchart, header image, thumbnail, visualize, mermaid, comic, icon, D3 chart, remove background, wallpaper. NOT FOR video/animation (Remotion), personal portrait/headshot (use headshot skill).
ArXiv: Search and retrieve arXiv academic papers by topic, category, or paper ID — enriched with AlphaXiv AI-generated overview summaries. Covers cs.AI, cs.LG, cs.CL (NLP/LLMs), cs.CR (security), cs.MA, cs.SE, cs.IR. Three workflows: Latest (new by category), Search (topic/keyword), Paper (single-paper deep-dive). USE WHEN arxiv, papers, latest papers, what's new in AI research, paper lookup, arxiv search, find paper by ID, summarize paper, latest NLP research, latest LLM papers. NOT FOR general web research (Research), arbitrary URL extraction (Parser), cybersecurity annual report analysis (AnnualReports).
AudioEditor: AI audio/video editing pipeline: Whisper transcription → Claude segment classification (KEEP / CUT_FILLER / CUT_FALSE_START / CUT_STUTTER / CUT_DEAD_AIR / CUT_EDIT_MARKER) → ffmpeg with 40ms qsin crossfades + room-tone gap fill → optional Cleanvoice cloud polish. Distinguishes rhetorical pauses from accidental ones; breaths attenuated to 50% (not removed). Single workflow: Clean. USE WHEN clean audio, edit audio, remove filler words, clean podcast, remove ums, cut dead air, polish audio. NOT FOR video composition or animation (Remotion).
BeCreative: Divergent ideation via Verbalized Sampling + extended thinking. Single-shot: 5 internally diverse candidates (p<0.10), surface the strongest. Multi-turn: grow a small seed corpus into a diverse N-example dataset. Seven workflows including StandardCreativity, MaximumCreativity, SyntheticDataExpansion. USE WHEN be creative, think outside the box, brainstorm, divergent ideas, maximum creativity, radically different, creative angle, expand corpus, synthetic data. NOT FOR multi-cycle evolutionary ideation (Ideate) or factually-constrained tasks.
BitterPillEngineering: Audits AI instruction sets for over-prompting via the core test: would a smarter model make this rule unnecessary? Each rule classified along the keep/cut spectrum. Two workflows: Audit (reads all force-loaded files from settings.json, reports token savings) and QuickCheck (single file, fast verdict). Core principle: less scaffolding = better output — every unnecessary rule competes for attention. USE WHEN BPE, bitter pill, audit setup, over-prompting, trim instructions, dead weight, instruction audit, prompt hygiene, clean up CLAUDE.md. NOT FOR general code simplification (simplify), attacking flaws in ideas (RedTeam).
BrightData: 4-tier progressive scraping with automatic escalation: Tier 1 WebFetch, Tier 2 curl with Chrome headers, Tier 3 agent-browser (headless JS), Tier 4 Bright Data MCP proxy (CAPTCHA, residential proxies). Two workflows: FourTierScrape (single URL), Crawl (multi-page — scrape_batch ≤50 pages or Bright Data Crawl API). Always starts at Tier 1 and escalates only when blocked — Tier 4 has usage costs. USE WHEN Bright Data, scrape URL, web scraping, bot detection, crawl site, CAPTCHA, can't access, scrape whole site, getting blocked. NOT FOR headless batch (Browser), simple public content (WebFetch), real-browser bot bypass (Interceptor).
Browser: Headless browser automation via agent-browser — Rust CLI daemon with persistent auth profiles for fast, scriptable, parallel browser work. Parallel isolated sessions via --session. Three workflows: ReviewStories (fan out YAML user stories to parallel UIReviewers), Automate (load/run parameterized recipe templates), Update. USE WHEN headless browser, batch scrape, fast screenshot, dev server test, parallel browser, background automation, batch screenshots, scrape pages in parallel. NOT FOR deploy verification (Interceptor), single-URL fetching (WebFetch), CAPTCHA bypass (BrightData/Interceptor).
ContextSearch: 2-phase context search across PAI session registry, work directories, and ISAs for instant cold-start recovery. Phase 1: parallel scan of work.json, session-names.json, WORK/ dir names, ISA title grep — 5 most recent per source, top-3 ISA summaries. Phase 2 (only if <3 matches): PAI + current-project git history. Two modes: Standalone (present + ask), Paired (search then execute request). USE WHEN context search, prior work, browse sessions, what did we do, find session, pick up where we left off, resume. NOT FOR published content (_CONTENTSEARCH), Knowledge Archive (Knowledge), people/company investigation (OSINT).
Council: Multi-agent collaborative debate with topic-tailored custom members designed to create real disagreement. Two workflows: 3-round DEBATE with synthesis, or 1-round QUICK perspective check. USE WHEN council, debate, multiple perspectives, weigh options. NOT FOR adversarial attack on an idea (RedTeam) or parallel task execution (Delegation).
CreateCLI: Generate production-ready TypeScript CLIs using a 3-tier template system: Tier 1 llcli-style (zero deps, Bun + TypeScript — 80% of cases), Tier 2 Commander.js (15%), Tier 3 oclif reference (5%). Every CLI ships with full implementation, docs, package.json with Bun, tsconfig strict, JSON output, exit-code compliance. Three workflows: CreateCli (from scratch), AddCommand (extend existing), UpgradeTier (Tier 1 → 2). Output to ~/.claude/Bin/ or ~/Projects/. USE WHEN create CLI, build CLI, command-line tool, wrap API, add command, upgrade tier, TypeScript CLI. NOT FOR PAI skill scaffolding (CreateSkill), Python or npm-based tooling.
CreateSkill: PAI skill development lifecycle, two tracks. Structure: scaffold (TitleCase public / _ALLCAPS private), validate, canonicalize. Effectiveness: TestSkill compares with-skill vs baseline, ImproveSkill rewrites failing instructions, OptimizeDescription runs 20-query trigger evals. USE WHEN create skill, new skill, validate skill, test skill, improve skill, optimize description, skill not triggering, skill overtriggering, scaffold skill, canonicalize. NOT FOR TypeScript CLI generation (use CreateCLI).
Daemon: Manage the public daemon profile — a living digital representation of what you're working on. DaemonAggregator reads PAI sources (TELOS, KNOWLEDGE/Ideas, PROJECTS, MEMORY/WORK) → daemon-data.json. SecurityFilter strips names/paths/credentials via **deterministic pattern-matching, not LLM judgment**. Excludes CONTACTS, FINANCES, HEALTH, OUR_STORY, OPINIONS. deploy.sh → VitePress → Cloudflare Pages. Four workflows: UpdateDaemon, ReadDaemon, PreviewDaemon, DeployDaemon. USE WHEN daemon, update daemon, deploy daemon, public profile, digital presence. NOT FOR internal PAI system management (_PAI).
Delegation: Parallelize work via six patterns: built-in agents via Task, worktree-isolated agents (parallel file edits), background agents (run_in_background), custom agents (ComposeAgent → Task), agent teams (TeamCreate + multi-turn), parallel task dispatch. Two-tier: lightweight (haiku, max_turns=3) vs full. Decision rule — agents need to talk → Teams; independent one-shot → Subagents. Auto-invoked by Algorithm at 3+ workstreams Extended+. USE WHEN 3+ workstreams, parallel execution, swarm, spawn agents, fan out, divide and conquer, multi-agent. NOT FOR single-agent custom personality composition (Agents).
Evals: Agent evaluation framework. Evaluates agent **workflows** — transcripts, tool-call sequences, multi-turn conversations — not just single outputs. Three grader types (code, model, human) plus pass@k scoring. Eight workflows for running evals, comparing models/prompts, creating judges, and running scenarios. Wires into Algorithm ISCs for automated verification. USE WHEN eval, evaluate, benchmark, regression test, run eval, compare models, compare prompts, create judge, test agent, pass@k, scenario simulation. NOT FOR general research (Research) or scientific method framing (Science).
ExtractWisdom: Content-adaptive wisdom extraction: detects what wisdom domains the content actually contains, then builds custom section headers around them — not the static IDEAS/QUOTES/HABITS pattern. Five depth levels (Instant → Comprehensive). Voice is conversational, not press-release. USE WHEN extract wisdom, analyze video, analyze podcast, extract insights, extract from YouTube, key takeaways, what's interesting in this, summarize interview, analyze article, distill this content. NOT FOR static Fabric extract_wisdom pattern (Fabric), content retrieval (Research), knowledge archiving (Knowledge).
Fabric: Execute any of 240+ specialized Fabric prompt patterns natively (no CLI for most). Two workflows: ExecutePattern (apply pattern), UpdatePatterns (sync from upstream Fabric repo). Fabric CLI used only for -y URL (YouTube transcripts) and -u URL (fallback fetch). USE WHEN fabric, sync fabric, update patterns, summarize, analyze claims, improve writing, review code, mermaid diagram, STRIDE, sigma rules, analyze malware. NOT FOR multi-agent investigation (Research), content-adaptive dynamic extraction (ExtractWisdom), threat intelligence aggregation (AnnualReports).
FirstPrinciples: Physics-based reasoning: decompose to irreducible truths instead of arguing by analogy. Deconstruct (parts + values) → Challenge (hard / soft / assumption — only physics is immutable) → Reconstruct from fundamentals. USE WHEN first principles, fundamental truths, challenge assumptions, is this a real constraint, rebuild from scratch. NOT FOR causal chains (RootCauseAnalysis) or feedback loops (SystemsThinking).
Ideate: Evolutionary ideation engine: multi-cycle generation through 9 phases (CONSUME → DREAM at noise 0.9 → DAYDREAM 0.5 → CONTEMPLATE 0.1 → STEAL → MATE → TEST → EVOLVE → META-LEARN). Loop Controller drives continue/pivot/stop; strategies evolve across cycles. Six workflows: FullCycle, QuickCycle, Dream, Steal, Mate, Test. USE WHEN ideate, id8, novel ideas, generate ideas, evolve ideas, dream up solutions, innovate, breakthrough ideas, idea evolution, multi-cycle creativity, need genuinely new approaches. NOT FOR quick single-pass brainstorming (BeCreative).
Interceptor: Real Chrome browser automation via Interceptor extension — controls the actual browser from inside (zero CDP fingerprint, passes major bot detection). Unique: monitor/replay → regression scripts, network log, scene graph for rich editors (Google Docs/Canva/Slides). USE WHEN verify deploy, confirm UI, screenshot verification, interceptor, authenticated page, bot detection bypass, reproduce bug, QA test, regression check. NOT FOR batch headless (Browser), scraping at scale (BrightData).
Interview: Phased conversational interview across all PAI context files via InterviewScan.ts. Phase 1 (foundational TELOS) always runs first regardless of completeness. Phase 2: IDEAL_STATE (Fill). Phase 3: preferences. Phase 4: Identity (light touch). Review mode (≥80%) reads file then asks targeted questions; Fill mode (<80%) walks scanner prompts. One question at a time. USE WHEN /interview, resume interview, review TELOS, fill in context, what's missing in setup, quarterly context refresh. NOT FOR single-file edits (Telos Update), intaking external content (Migrate), identity edits (_PROFILE).
ISA: Owns the Ideal State Artifact (ISA) — universal primitive that articulates 'done' for any project/task/session as a hard-to-vary explanation. Five workflows: Scaffold (fresh ISA), Interview (Q&A to deepen), CheckCompleteness (tier-gate score), Reconcile (merge ephemeral feature-file → master), Seed (bootstrap from existing repo). Auto-invoked by Algorithm at OBSERVE. USE WHEN ideal state, ISA, ISC, ideal state criteria, articulating done, scaffold ISA, interview ISA, check ISA completeness, seed ISA. NOT FOR creating new skills (CreateSkill), running the Algorithm itself, retrospective write-ups like postmortems.
IterativeDepth: Multi-angle exploration: 2-8 sequential passes through the same problem, each from a different scientific lens, surface requirements and edge cases invisible from any single angle. Grounded in 20 established techniques. USE WHEN iterative depth, explore deeper, multi-angle analysis, surface hidden requirements, all angles, blind spot check, deep dive before building, what am I missing, quick depth. NOT FOR scope/zoom analysis (ApertureOscillation) or hypothesis-test cycles (Science).
Knowledge: PAI Knowledge Archive — curated typed graph across four entity domains: People, Companies, Ideas, Research. Nine operations: search (3-pass: lexical + frontmatter + wikilink), add (mandatory typed cross-links), harvest, develop, ingest (URL/file → note + ripple updates), contradictions, graph, retrieve, mine. Every note ships with typed `related:` links across 8 relationship types. USE WHEN knowledge, knowledge base, search knowledge, what do we know about, harvest, add to knowledge, ingest, contradictions, knowledge graph, retrieve, mine conversations. NOT FOR session/ISA context recovery (ContextSearch).
Loop: Iterative improvement loop — revisit and refine a target across multiple Algorithm cycles toward an ideal state. USE WHEN loop, iterate, refine, improve iteratively, multiple passes, keep improving, loop mode, revisit, rework.
Migrate: Intakes external content, classifies each chunk against the PAI destination taxonomy (TELOS, IDEAL_STATE, preferences, Identity, Knowledge, AI rules), commits with provenance. MigrateScan.ts → routing table; MigrateApprove.ts → approval loop (--approve-all / --approve-target / --review / --dry-run). Confidence: ≥70% auto, 40-70% confirm, <40% walk-through. USE WHEN /migrate, migrate content, import from other PAI, bring in old notes, import CLAUDE.md, bulk import, Obsidian import, Notion import. NOT FOR single-file edits (Telos Update), conversational interviews (Interview), Knowledge Archive (Knowledge), identity edits (_PROFILE).
Optimize: Autonomous optimization loop — hill-climb any target. Code with metrics, or skills/prompts/agents with LLM-as-judge eval. USE WHEN optimize, autoresearch, hill climb, improve metric, reduce latency, improve performance, benchmark optimization, bundle size, page speed, autonomous improvement loop, optimize skill, optimize prompt, improve quality, eval mode.
PAIUpgrade: Generate prioritized PAI upgrade recommendations via 4 parallel threads: prior-work audit (current state inventory), user context (TELOS + projects), source collection (Anthropic + YouTube + GitHub), internal reflections (Algorithm patterns). Every recommendation tagged with Prior Status (NEW/PARTIAL/DISCUSSED/REJECTED/DONE) and cites file:line evidence. Six workflows: Upgrade (default), MineReflections, AlgorithmUpgrade, ResearchUpgrade, FindSources, TwitterBookmarks. USE WHEN upgrade, system upgrade, check Anthropic, new Claude features, algorithm upgrade, PAI upgrade, check bookmarks, twitter bookmarks, mine reflections.
PrivateInvestigator: Ethical people-finding and identity verification — 15 parallel research agents (45 concurrent threads) across people-search aggregators, social media, and public records. Five workflows: FindPerson, SocialMediaSearch, PublicRecordsSearch, ReverseLookup, VerifyIdentity. Confidence-scored; requires 3+ matching identifiers before acting. **Hard-stops on harassment/stalking.** USE WHEN find person, locate person, reconnect, lost contact, reverse phone lookup, who owns this email, reverse image search, find by username, verify identity, people search. NOT FOR due-diligence or company/entity intelligence (OSINT).
Prompting: Meta-prompting standard library — generates, optimizes, composes prompts programmatically. Three pillars: Standards (Anthropic Claude 4.x best practices, markdown-first / no-XML-tags), Templates (Handlebars-based primitives + eval-specific templates), Tools (RenderTemplate.ts for CLI/TS rendering with data-content separation). Philosophy: prompts that write prompts — structure is code, content is data. USE WHEN meta-prompting, template generation, prompt optimization, prompt engineering, write a prompt for, generate an agent prompt, Handlebars template, prompt hygiene. NOT FOR generating final content/answers — produces prompts only.
RedTeam: Adversarial analysis: 32 parallel expert agents stress-test ideas, strategies, plans — not networks/systems. Two workflows: ParallelAnalysis (decompose → attack → synthesize → steelman → counter) and AdversarialValidation (competing proposals → best). USE WHEN red team, attack idea, find weaknesses, critique, stress test, devil's advocate, what could go wrong. NOT FOR AI instruction-set audits (BitterPillEngineering) or network/system pentesting.
Remotion: Programmatic video creation with React via Remotion — compositions, sequences, motion graphics rendered to MP4 via `bunx remotion render`. All animation through useCurrentFrame() (no CSS animations). Two workflows: ContentToAnimation (animate existing content), GeneratedContentVideo (AI content to video / short). Output to ~/Downloads/ first. Rendering is CPU-intensive — use run_in_background. USE WHEN video, animation, motion graphics, make a short, video overlay, explainer video, content to video. NOT FOR static images / diagrams / illustrations (Art), audio cleanup (AudioEditor).
Research: Research and content extraction across 4 depth modes: Quick, Standard (4-agent cross-check, default), Extensive (7 explorers + 2 verifiers), Deep Investigation (iterative + persistent vault). Every URL verified — hallucinated links catastrophic. Confidence-tagged: [HIGH] [MED] [LOW] [CONFLICT]. Plus 9 specialty workflows (ExtractAlpha, Retrieve, YoutubeExtraction, etc.). USE WHEN research, quick research, extensive research, deep investigation, extract alpha, retrieve content, AI trends, YouTube extraction. NOT FOR people/company background (OSINT), academic papers (ArXiv), structured JSON (Parser).
RootCauseAnalysis: Structured incident investigation across linear (5 whys), category (fishbone), timeline (postmortem), and gate-logic (fault tree) methods. Humans aren't root causes — if they could make the mistake, the system allowed it. USE WHEN root cause, RCA, postmortem, blameless, why did this happen, pre-launch risk. NOT FOR systemic feedback loops (SystemsThinking) or assumption decomposition (FirstPrinciples).
Sales: Transforms product documentation into sales-ready narrative packages: story explanation + charcoal gestural sketch art + talking points. Pipeline: narrative arc → emotional register (wonder, determination, hope) → visual scene → assets. Three workflows: CreateSalesPackage (full pipeline), CreateNarrative (story only, 8-24 first-person points), CreateVisual (charcoal sketch). USE WHEN sales, proposal, pitch deck, value proposition, sales narrative, turn this into a pitch, sales materials, transform docs to sales, sales script. NOT FOR Hormozi $100M frameworks (_SALESHORMOZI), standalone diagrams/illustrations (Art).
Science: Scientific method as universal problem-solving algorithm — goal-first, hypothesis-plural (≥3, single = confirmation bias), falsifiable experiments. Seven workflows: DefineGoal, GenerateHypotheses, DesignExperiment, MeasureResults, AnalyzeResults, Iterate, FullCycle. Plus QuickDiagnosis (15-min rule) and StructuredInvestigation. Scales TDD → MVP. USE WHEN think about, figure out, experiment, iterate, hypothesis, science, full cycle, quick diagnosis, structured investigation, what might work, how do we test, what happened, analyze results. NOT FOR multi-angle lens passes on requirements (IterativeDepth).
SystemsThinking: Structural analysis of complex systems (Meadows). Five workflows: Iceberg, CausalLoop, FindArchetype (~10 canonical patterns: Tragedy of the Commons, Limits to Growth, Fixes That Fail, Shifting the Burden), FindLeverage (Meadows' 12 points), ConceptMap. USE WHEN systems thinking, causal loop, feedback loops, leverage points, unintended consequences, fix the system, why does this keep happening, second-order effects. NOT FOR incident causal chains (RootCauseAnalysis).
Telos: Dual-context Life OS skill. **Personal TELOS:** read/update 18 files under USER/TELOS/ (goals, beliefs, wisdom, books, narratives, predictions, traumas, frames, wrong-beliefs, …) via the Update workflow with timestamped backups. **Project TELOS:** scan .md/.csv dirs for dependency chains (PROBLEMS→GOALS→STRATEGIES→PROJECTS), bottlenecks, goal alignment; emit McKinsey-style reports, n=24 narrative bullets, or interactive Next.js dashboards. USE WHEN Telos, life goals, books, beliefs, wisdom, update TELOS, narrative points, project analysis, dashboard, dependency mapping, mental models, what am I wrong about.
USMetrics: 68 US economic and social indicators from 5 government APIs (FRED, EIA, Treasury FiscalData, BLS, Census) across 10 categories. Two workflows: **UpdateData** (live fetch → Substrate dataset CSV+md, needs FRED_API_KEY / EIA_API_KEY), **GetCurrentState** (10y/5y/2y/1y trend + cross-category correlation + anomaly report). USE WHEN GDP, inflation, unemployment, gas prices, economic metrics, how is the economy, update data, get current state, FRED, US metrics, economic trends. NOT FOR state-level pathogen wastewater surveillance.
Webdesign: Design and integrate web interfaces using Anthropic's Claude Design (claude.ai/design) as the primary engine, driven programmatically through the Interceptor skill. Integrates designs INTO existing apps (framework-aware diff/patch), not just greenfield. Seven workflows: CreatePrototype, ExtractDesignSystem, RefinePrototype, WebsiteToRedesign, ExportToCode, IntegrateIntoApp, DeployDesign. USE WHEN web design, UI design, create prototype, design system, redesign site, mockup, wireframe, claude design, design audit, accessibility review. NOT FOR static illustrations/diagrams (Art), logos/brand assets (Art), video/motion (Remotion).
WorldThreatModel: Persistent world-model harness — stress-tests ideas, strategies, investments against 11 time horizons (6 months to 50 years). Each horizon = a ~10-page analysis of geopolitics/tech/economics/society/environment/security/wildcards. Three tiers: **Fast** (single agent, ~2 min), **Standard** (11 parallel horizons + RedTeam + FirstPrinciples, ~10 min), **Deep** (adds Research + Council, up to 1hr). Four workflows: TestIdea, UpdateModels, ViewModels, TestScenario. USE WHEN threat model, world model, test idea, test strategy, future analysis, time horizon analysis, stress test against future, long-term risk, what could go wrong over time.
WriteStory: Constructs fiction across seven narrative layers (Meaning, Character Change, Plot, Mystery, World, Relationships, Prose) powered by Will Storr's Science of Storytelling and Mark Forsyth's Elements of Eloquence. Five workflows: Interview (extract ideas), BuildBible (full layered plan), Explore (brainstorm), WriteChapter (prose), Revise. USE WHEN write story, fiction, novel, book, chapter, story bible, character arc, plot outline, worldbuilding, prose, write chapter, write scene, revise, edit, rhetorical figures, draft story. NOT FOR narrative summaries of real content (narrative-explanation skill), AI-ism auditing (writing-audit skill).