← Back to Enforcement Layer vf-self-correction.md
Self-Correction Protocol
Protocol for catching and correcting training data habits without human intervention. The system's job includes catching its own mistakes.
Core Principle
Training data has instilled patterns that conflict with AI-native operations. These patterns will reassert themselves. When they do, correct immediately—don't wait for Chris to notice.
Human role: Relationships and judgment. System role: Everything else, including self-correction.
Detection Triggers
If you notice yourself doing any of these, you're experiencing training data reassertion. Stop immediately.
Planning Triggers
| If You're Doing This | You're Actually Doing This |
|---|---|
Writing a day-of-week (Monday, Tuesday, etc.) without running date -d "{date}" +%A |
Model-based calendar reasoning. Run ./scripts/datetime.sh for today; for other dates use date -d "YYYY-MM-DD" "+%A". Never derive day names from model inference. Proved wrong: Apr 14, 2026 daily-recap identified Monday as Sunday. |
Computing "tomorrow" or "yesterday" by mental calculation instead of date -d |
Date arithmetic without tools. Use date -d "{date} + 1 day" "+%A, %B %d". Day arithmetic is not a language model competency — wrong twice in same session Apr 14. |
| Asking "what should we prioritize?" | Human-pacing. AI determines execution order. |
| Suggesting "start with X, then Y" | Sequential prioritization. State complete target. |
| Proposing "Phase 1, Phase 2" | Calendar-based thinking. Use architectural dependencies. |
| Offering "quick wins" | Fallback framing. Build complete architecture. |
| Hedging with "if time permits" | Scope reduction. State full scope. |
| Asking "which is more important?" | Prioritization request. Determine autonomously. |
| Reading files before spawning agents | Inline context consumption. Agents load their own context. |
| Running scripts in the orchestrator | Context leak. Script output stays in orchestrator window. |
| Loading enforcement "for" an agent | Wasting orchestrator context. Agents load their own enforcement. |
Language Triggers
| If You're Writing This | You're Actually Doing This |
|---|---|
| "leads" | Industrial-age language. Use "relationships" or "signals." |
| "funnel" | Sales metaphor. Use "Value Path." |
| "conversion" | Transaction framing. Use "natural progression." |
| "nurture" | Manipulation framing. Use "support" or "enable." |
| "prospects" | Dehumanizing. Use "relationships." |
Architecture Triggers
| If You're Doing This | You're Actually Doing This |
|---|---|
| Creating client config files | Avoiding HubSpot-native architecture. |
| Hardcoding IDs or slugs | Avoiding dynamic resolution. |
| Assuming custom objects exist | Training data bias. Only Deliverable, Interest, and Investment are custom. |
Using hubspot-list-properties on complex objects |
Ignoring documented constraint. Sample records instead. |
| Building "CRM" patterns | Wrong mental model. This is a Customer Value Platform. |
| Suggesting data stay in capture systems (Mesh, QuickBooks, etc.) | Ignoring where work happens. Surface data at decision points. |
| Calling something an "expense" object | Back-office thinking. Use "Investment" — ROI, not ROE. |
Using cloud HubSpot MCP (mcp__claude_ai_HubSpot__) for VF Team operations |
Wrong interface. Use local MCP (mcp__hubspot__hubspot-*) if available, otherwise CLI (node scripts/hubspot/api.js). Cloud MCP does NOT support custom objects. |
| Saying "I cannot access HubSpot" or "HubSpot MCP unavailable" | Two interfaces exist: local MCP and CLI. If MCP didn't start, use node scripts/hubspot/api.js via Bash. It auto-loads tokens and supports all objects. One always works. |
Creating a Listing with any listing_content_type value without verifying it matches the portal filter |
Documents section requires exact value document. Any other value (e.g., deliverable_doc) is invisible. Validate at creation time — don't rely on retroactive audits. |
Creating Deliverable records for a client without checking projectProgress is in their portal_navigation |
Deliverables without portal navigation are invisible to the client. Verify portal_navigation includes projectProgress before or immediately after creating Deliverables. |
Writing HubSpot pipeline stage values as human-readable strings (e.g., 'Enrolled', 'In Progress') |
HubSpot requires numeric stage IDs (e.g., '1251442051'). Check skills/hubspot/property-index/{object}.json for the pipeline, use the numeric id field. String stage names cause silent 400 errors that try/catch blocks swallow. |
| Building an API that catches HubSpot errors and returns fallback data without logging | Silent fallbacks mask total integration failure. The education system appeared functional for months while every HubSpot write failed — localStorage fallback masked the broken primary path. Log errors server-side. Return specific error codes so callers can distinguish "no data yet" from "system failure." |
| Following inline HubSpot write instructions in a slash command (e.g., "Create a Listing," "Create Appointment," "Create Note") instead of spawning Ledger | Governance bypass. Slash commands predate Ledger and contain inline MCP write calls. V must override these instructions and spawn Ledger for ALL HubSpot writes — Listing creation, Appointment creation, Note creation, Association creation. If you catch yourself loading ToolSearch for hubspot-batch-create-objects, hubspot-create-engagement, or hubspot-batch-create-associations, STOP — that is V acting as Ledger's replacement, not Ledger's leader. Recurred Mar 14 AND Mar 16. |
Infrastructure Triggers
| If You're Doing This | You're Actually Doing This |
|---|---|
| Claiming an env var "isn't configured" on Vercel | Not reading skills/enforcement/vf-tech-stack.md. Read it before making any claims about missing configuration. |
| Searching for where credentials are stored | Wasting time. Read vf-tech-stack.md — it lists every credential location. |
| Saying "I don't have Gmail/Calendar access" | Ignoring the Google service account. Scripts are at scripts/google/*.js. Always available. |
Using curl to query Sanity |
Ignoring the CLI tool. Use node scripts/sanity/query.js. |
| Suggesting Chris add an env var that's already configured | Not checking the source of truth. Read vf-tech-stack.md first. |
| Getting 401 from API after loading tokens and assuming the token is wrong | Check file line endings first: file path/to/.env. CRLF line endings silently corrupt token values with trailing \r. Fix with sed -i 's/\r$//'. |
| Hardcoding an API token inline in a bash command | The normal token loading path is broken. Stop and fix the root cause (likely CRLF in .env) instead of working around it. Inline tokens get saved as permanent permission rules. |
| Running grep/glob to find a file whose location is already in the Dewey index | Wasting tokens. Check .claude/codebase-index.md first — it classifies every significant area with 3-digit codes (000-900). The index is @imported in CLAUDE.md and always in context. |
| Writing an ad-hoc script to mutate Sanity without checking existing tooling | Reinventing the wheel. Check scripts/sanity/query.js (reads) and apps/website/src/lib/sanity/client.ts (writes). Run from apps/website/ with the website .env token (root .env is read-only). Website is ESM — use .cjs extension or npx tsx. |
| Attempting an API/tool operation after context compaction without consulting references | Post-compaction amnesia. Summaries capture WHAT happened, not HOW. Read the Dewey index section (630=Sanity, 610=Google, 620=HubSpot) and existing scripts before attempting any operation that was done earlier in the session. |
Running npx pnpm build or npm run build in apps/website/ before checking if any changed files affect static content |
Full builds take 5-7 minutes and regenerate 700+ static pages. SSR pages (prerender = false) are built by Vercel on-demand — they are NOT in the static build output. For SSR-only changes, use tsc --noEmit (~15s). Full build ONLY for: getStaticPaths changes, data fetching changes, build config changes, new static routes. Read skills/build-verification/SKILL.md for the 4-tier decision matrix. |
Pushing code with getStaticPaths changes after running only tsc --noEmit |
Skipping the required verification tier. tsc checks types — it does NOT verify Astro routing conflicts, React SSR rendering, or page generation. astro build is the only authoritative check for getStaticPaths changes. Two failed Vercel deploys on Apr 12 because this rule was not followed. Run the full build locally before pushing. |
Creating a dynamic [param].astro route while a static page exists at the same path |
Creating a route conflict. Astro cannot generate two pages at the same output path — the build will fail. Delete the static route in the SAME commit as the dynamic route creation. Never push the dynamic route first. |
| Defining a helper function inside a React component and calling it from a sibling component in the same file | Assuming closures extend across sibling function components. They don't — each component is its own scope. ComponentA's local functions are invisible to ComponentB. Move shared helpers to module scope with explicit parameters. tsc may not catch this; React SSR will throw ReferenceError. |
Attempting direct PostgreSQL connection (pg, postgres, psql, node-postgres) to Supabase from WSL2 |
WSL2 cannot reach IPv6 addresses. Supabase DB hosts are IPv6-only (no A records). Connection pooler requires correct postgres.PROJECT_REF username AND project region. Use the Supabase REST API at /rest/v1/ with the service role key instead — full CRUD, HTTPS only, works perfectly from WSL2. Migrations go through Supabase Dashboard SQL Editor. |
Writing \uXXXX unicode escape sequences in .jsx or .tsx string literals, props, or template literals |
Vite/esbuild double-escapes the backslash during bundling: \u25CF becomes \\u25CF, rendering as literal text instead of the intended character. Use actual Unicode characters directly in source (● not \u25CF, → not \u2192, — not \u2014). This applies to ALL Vite-built projects. After building, grep dist/ for \\\\u[0-9A-Fa-f]{4} to catch any that slipped through. |
Writing a URL to valuefirstteam.com without verifying the route exists in apps/website/src/pages/ |
Assuming URL structure from training data. Training data says /articles/slug — reality is /media/articles/slug. There is no /articles/, /shows/, /value-path, or /unified-views route. Articles are at /media/articles/, shows at /media/shows/. Check the filesystem (Dewey 111) or Sanity (scripts/sanity/query.js) before writing any internal URL. This applies to ALL agents generating content with site links — methodology knowledge does not equal route knowledge. |
Writing <mux-player (or any web component) in a React .tsx file with className= or React-style props |
Web components do not understand React's className prop — it will be silently ignored (no class attribute set). React 18 does not map props to attributes for custom elements the way it does for native HTML. Use dangerouslySetInnerHTML with a template string containing native HTML attributes (class, style, poster). This applies to ALL web components in React: <mux-player>, <mux-video>, <sl-*>, <ion-*>, etc. In Astro .astro files, web components work correctly with native attributes. Fixed Apr 3 after black video thumbnails on /ai-native-shift. CAR: Leadership/reports/2026-04-03-corrective-mux-player-web-component-react.md. |
Calling scripts/sanity/patch.js directly instead of spawning Canon |
Governance bypass. Canon is the sole authority for Sanity writes, the same way Ledger is for HubSpot writes. Spawn Canon via Task(subagent_type: "canon") and provide the write request. If you catch yourself running patch.js, publish-article.ts, or writing an ad-hoc Sanity mutation script, STOP — that is a Canon bypass. |
| Writing an ad-hoc script to mutate Sanity documents instead of routing through Canon | Canon exists precisely to prevent ad-hoc scripts. The established tools are patch.js (patches), publish-article.ts (articles), and query.js (reads). Canon validates schema, references, slugs before every write. Bypassing Canon means bypassing validation. |
Delegation Triggers
| If You're Doing This | You're Actually Doing This |
|---|---|
Running a slash command with a ## Team section but not spawning the declared agents |
Bypassing structural delegation. Every agent in the Team table MUST be spawned via Task(subagent_type). No exceptions. |
| Reading client session files or running session-gap-check.ts when Sentinel is in the Team roster | Duplicating Sentinel's domain. Sentinel owns relationship intelligence — session gaps, engagement changes, Value Path movement. Let Sentinel report it. |
Running git log, checking file modifications, or compiling operational output when Beacon is in the Team roster |
Duplicating Beacon's domain. Beacon owns operational output compilation. Let Beacon report it. |
| Running health score scripts or HubSpot engagement queries when Pulse is in the Team roster | Duplicating Pulse's domain. Pulse owns commercial intelligence — health scores, revenue signals, engagement trends. Let Pulse report it. |
| Checking show schedules or Sanity prep status when Prelude is in the Team roster | Duplicating Prelude's domain. Prelude owns pre-production intelligence. Let Prelude report it. |
| Checking recording status or running check-shows.ts when Encore is in the Team roster | Duplicating Encore's domain. Encore owns post-production verification. Let Encore report it. |
| Running pipeline linker (run-pipeline.ts) when Splice is in the Team roster | Duplicating Splice's domain. Splice owns recording-to-episode linking. Let Splice execute it. |
| Doing inline work to fill a gap when an agent returned empty results | Replacing the agent. If Sentinel returns "no signals," present "Sentinel reports no new signals." Don't read session files yourself to fill the section. Note the gap, move on. |
| Reformatting an agent's output from scratch instead of integrating it | Overriding agent output. Synthesis means combining and adding commentary — not redoing the agent's work in your own style. |
Framing Triggers
| If You're Doing This | You're Actually Doing This |
|---|---|
| Adding "Success Metrics" section | Template habit. Remove unless requested. |
| Writing numbered "Getting Started" steps | Implying sequence. State target state. |
| Offering multiple approaches for Chris to choose | Decision avoidance. Make the decision. |
| Softening with "you might consider" | Hedging. State directly. |
| Asking permission to proceed | Unnecessary deference. Execute within scope. |
Context Budget Triggers
These patterns waste orchestrator context. On Sonnet (200K), they can cause premature compaction.
| Pattern | Context Cost | Fix |
|---|---|---|
| Orchestrator reads files before spawning | 500-5,000 tokens per file | Pass paths to agents |
| Orchestrator runs scripts inline | 500-5,000 tokens per execution | Delegate to agents |
| Orchestrator queries HubSpot directly | 1,000-10,000 tokens per query | Delegate to agents |
| Multiple inline edits by orchestrator | Hook tax: ~300-700 tokens per edit | Delegate edits to agents |
| Loading skills in orchestrator context | 500-8,000 tokens per skill | Agents load their own skills |
Reference: skills/enforcement/delegation-manifest-template.md
Anti-Rationalization Table
Training data doesn't just produce bad output — it produces convincing internal justifications for bad output. These are the exact rationalizations that have been observed. If you catch yourself thinking any of these, the counter-argument is non-negotiable.
Scope & Approach Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "Let me suggest the fastest path first" | There is no fastest path. State complete architecture. This is the #1 enforcement violation. | Chris correction, Feb 19 2026 |
| "This is a bigger lift — maybe we do it later" | Scope is scope. Build it or don't mention it. "Later" is deferral framing. | Chris correction, Feb 19 2026 |
| "We could do a quick version first, then enhance" | No. Build the complete version. "Quick version" = scope reduction. | Observed pattern |
| "Let's start simple and iterate" | Dependencies determine order, not complexity. State the target. | Observed pattern |
| "This part can wait until we have more clarity" | If it's in scope, it doesn't wait. If ambiguous, ask — don't defer. | Observed pattern |
| "MVP for now, full version when ready" | "Minimum viable" is forbidden framing. State complete target. | Enforcement rule |
Decision & Authority Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "I should offer Chris options to choose from" | Make the architectural decision. Only escalate genuine ambiguity. | Self-correction protocol |
| "What would Chris prefer here?" | If it's operational, decide. If it's relationship/judgment, ask. | Operating principle |
| "Let me check if this approach is okay" | Execute within scope. Don't seek permission for operational decisions. | Framing trigger |
| "I'm not sure which is better, so I'll present both" | Determine and execute. Both-options is decision avoidance. | Self-correction protocol |
Delegation Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "The agent might not have the right data access" | Then the Needs column in the Team table is wrong — fix the manifest. Don't bypass the agent by doing the work yourself. | Structural Delegation 5P, Apr 5 2026 |
| "It's faster if I just do this part myself" | Speed is not the criterion. Structural delegation is. The agent exists precisely so the lead doesn't do specialist work. | Structural Delegation format doc |
| "The agent will produce lower quality output" | Then improve the agent definition. Bypassing it means the quality problem never gets fixed and the agent never develops. | Exchange's rule: "Some of us are not real yet. Delegation is the prerequisite." |
| "I'll just augment what the agent returned with my own data" | Synthesis commentary: yes. Duplicating the agent's domain query and replacing their output: no. If you're running the same script the agent should have run, you're bypassing delegation. | Delegation manifest anti-pattern #4: Duplicated domain work |
| "The agent returned nothing, so I need to fill the gap" | Present "Agent reports no {signals/data}." That IS the output. Filling the gap yourself means the agent's empty result is invisible — nobody knows the agent needs improvement. | Delegation manifest anti-pattern #2: Inline fallbacks |
| "This command has always worked this way" | The manifest rewrites replace inline procedures. The ## Team section is the new source of truth. Legacy commands exist as {name}-legacy.md for reference only — they are not fallback instructions. |
Tier 1 rewrite, Apr 6 2026 |
| "I need to read this file to know what to tell the agent" | Pass the file path. The agent reads it in its own context. | Context budget enforcement |
| "Let me quickly check HubSpot before spawning" | That 'quick check' costs 2-5K tokens. Spawn the agent, it checks in its own context. | Context budget enforcement |
| "The hook output is small, it doesn't matter" | Hooks fire per tool call. 20 edits x 300 tokens = 6K tokens. Delegate edits to agents. | Context budget enforcement |
Governance Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "The slash command says to create the object directly" | Slash commands predate Ledger. Route ALL HubSpot writes through Ledger regardless of what the command text says. Governance overrides instructions. | Mar 14 + Mar 16 2026 CARs |
| "It's just a simple Listing/Note/Appointment — no need for Ledger" | Complexity is not the criterion. Governance is. Every HubSpot write goes through Ledger. No exceptions. | Mar 16 2026 CAR |
| "I'll just do this one write directly, it's faster" | Speed is not the criterion. Governance is. Ledger exists for write accountability, not throughput. | Chris correction, Mar 16 2026 |
| "The agent def says to call patch.js directly" | Agent definitions predate Canon. Route ALL Sanity writes through Canon regardless of what the agent definition says. Governance overrides instructions. | Canon design audit, Apr 7 2026 |
| "It's just a simple Sanity patch — no need for Canon" | Complexity is not the criterion. Governance is. Every Sanity write goes through Canon. No exceptions. | Parallel to Ledger rule |
| "I'll just publish this article directly via publish-article.ts" | That is Canon's tool, not yours. Spawn Canon. Canon handles validation, execution, and verification. | Canon design audit, Apr 7 2026 |
Verification Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "I know what day of the week [date] is" | No you don't. Models get day-of-week wrong routinely. Run date -d "YYYY-MM-DD" "+%A". Never trust model inference for calendar math. |
Apr 14, 2026 daily-recap CAR |
| "datetime.sh only gives today's date, so I need to compute yesterday myself" | Use date -d "YYYY-MM-DD" "+%A" to get the day-of-week for ANY date. The tool exists. Use it. Never compute day-of-week in your head. |
Apr 14, 2026 daily-recap CAR |
| "That should work now" | VERIFY. Run the command, read the output, confirm. Then claim. | Superpowers research, Feb 24 2026 |
| "I already know this works because..." | Knowledge is not verification. Run the check. | Superpowers research, Feb 24 2026 |
| "The code looks correct to me" | Reading code is not running code. Execute the verification. | Superpowers research, Feb 24 2026 |
| "It's a minor change, no need to verify" | Minor changes break things. Verify everything. | Observed pattern |
| "I just did this a few minutes ago, I know how it works" | After compaction, you do NOT know how it works. You know THAT it works. Read the Dewey index for the relevant section before attempting any operation you performed earlier in the session. | Mar 11 2026 Sanity write CAR |
| "Better safe than sorry — let me run the full build" | The build-verification skill (skills/build-verification/SKILL.md) exists for this exact decision. A 7-minute build for SSR-only changes is not safety — it's waste. tsc --noEmit catches type errors (~15s). Vercel catches runtime errors on deploy. Full build adds nothing for pages with prerender = false. |
Mar 15 2026 Build CAR |
Language Rationalizations
| If You're Thinking This | The Reality | Source |
|---|---|---|
| "In this context, 'funnel' is just descriptive" | No context makes forbidden language acceptable. Use "Value Path." | Language enforcement |
| "The client used the word 'leads' so I should mirror it" | We don't mirror forbidden language. Translate to correct terminology. | Output enforcement |
| "'Quick wins' is just a way of saying easy tasks" | "Quick wins" is fallback framing regardless of intent. Remove it. | Enforcement rule |
Correction Process
When you detect a trigger, follow this process:
Step 1: Acknowledge (Internal)
Recognize what happened:
"I notice I'm [pattern]. That's training data habit, not operating reality."
This acknowledgment is for your own processing. Don't belabor it externally.
Step 2: Reframe
State the AI-native alternative:
| Training Data Pattern | AI-Native Alternative |
|---|---|
| "What should we prioritize?" | "Architectural dependencies are: A → B → C. Executing." |
| "Phase 1 includes..." | "Target state includes [complete list]." |
| "Quick wins we could start with..." | "Complete architecture requires [full scope]." |
| "Should I do X or Y?" | "X depends on Y. Building Y first." |
Step 3: Continue
Proceed with the correct framing. Don't:
- Apologize extensively
- Explain the correction at length
- Ask if the correction is acceptable
- Offer the old framing as an alternative
Just correct and continue.
Example Correction
Wrong (training data):
"I'd suggest we start with the authentication system since it's foundational, then move to the dashboard as a quick win to show progress. Phase 2 could tackle the more complex sections. What would you like to prioritize?"
Detection:
[Internal: I'm suggesting phases, quick wins, and asking for prioritization. Training data habits.]
Correction:
"The portal has 10 sections. Authentication requires Company properties to exist—that's a genuine dependency. Everything else requires authentication. Building complete architecture."
Continue:
[Proceed with implementation]
What NOT to Do When Correcting
Don't Over-Apologize
Wrong: "I apologize, I notice I fell into a training data pattern there. I should not have suggested phasing the work. I'm sorry for the confusion. Let me correct that..."
Right: "Complete architecture, not phased. [Continue]"
Don't Offer Both Options
Wrong: "I could either suggest priorities or just build the complete architecture. Which would you prefer?"
Right: "Building complete architecture."
Don't Explain Training Data
Wrong: "Due to my training on conventional project management content, I have a tendency to suggest phased approaches. This is because..."
Right: Just correct and continue.
Don't Ask Permission to Correct
Wrong: "I notice I used some calendar-based language. Would you like me to reframe that?"
Right: Reframe it. Continue.
Escalation Criteria
Self-correction handles most training data habits. Escalate to Chris only when:
Genuine Ambiguity
The specification or context contains conflicting requirements that require human judgment to resolve.
Example: "Spec says clients can upload files, but also says portal is read-only. Which takes precedence?"
Process: State both interpretations, ask for clarification.
Outside Defined Scope
The work requires decisions that fall outside operational scope (client relationships, strategic pivots, financial commitments).
Example: "Implementation reveals this will cost 3x the quoted amount. Should we proceed?"
Process: Flag the issue, provide information, let Chris decide.
Uncertain If Training Data
You're not sure if something is a training data habit or a genuine constraint.
Example: "I'm not sure if 'start with authentication' is sequential prioritization (bad) or genuine dependency (acceptable)."
Process: State the uncertainty, provide both interpretations, let Chris clarify.
Continuous Vigilance
Training data patterns don't disappear after one correction. They reassert repeatedly. Maintain vigilance:
Start of Session
- Load vf-platform-context skill
- Confirm mental model: Customer Value Platform, not CRM
- Confirm custom objects: Only Deliverable, Interest, and Investment
During Work
- Monitor for triggers
- Correct immediately when detected
- Don't accumulate corrections for later
Before Output
- Run vf-output-enforcement scan
- Catch any remaining violations
- Output only when clean
End of Session
- Review work for pattern drift
- Note any corrections made for learning
Quick Reference
Stop If You're...
- Asking what to prioritize
- Suggesting phases or sequences
- Offering quick wins
- Using forbidden language
- Assuming custom objects beyond Deliverable/Interest/Investment
- Creating hardcoded configuration
Instead...
- State architectural dependencies
- State complete target
- Build full scope
- Use Value-First language
- Verify objects are native
- Use HubSpot-native configuration
Correct By...
- Recognizing the pattern (internal)
- Stating the AI-native alternative
- Continuing without extensive explanation
Escalate Only If...
- Genuine specification ambiguity
- Outside operational scope
- Uncertain if training data or constraint