Training copy - internal field guide

Zant's Claude Code Field Guide

Everything wired into this machine's Claude Code setup: the ground rules in CLAUDE.md, the two workflow axes, all 71 skills with what they do and how to invoke them, and the memory system underneath. Read top to bottom to learn the system, or drill into any skill card when you need it.

Source of truth: ~/CLAUDE.md + ~/.claude/skills/ + the mattpocock-skills plugin. Snapshot: 2026-08-06.

1. The mental model: three systems, one stack

The whole setup reduces to three independent systems that compose. If you remember nothing else, remember this split - every skill below belongs to one of them.

Vertical axis

How one piece of work ships

The mattpocock-skills plugin: sharpen the idea, spec it, split it into tickets, implement each one test-first, close with review. One flow, idea to shipped code.

Horizontal axis

Whether work splits across workers

/so plus the Solo MCP: a triage agent decides pass-through, "grill me first", or a multi-agent board with todos and scratchpads. Always explicit opt-in.

Memory layer

Where the "why" survives

Persistent memory files + Open Brain (semantic store). Decisions, corrections, and project state outlive any one conversation, whichever axis produced them.

The rule that keeps them from fighting: each altitude of work has exactly one owner. A tracker (if the repo has one) owns specs and tickets; Solo owns one agent's in-flight mechanics; memory records outcomes. Section 4 covers how a repo declares which applies.

2. Ground rules (the CLAUDE.md contract)

These load into every session. They are behavioral guardrails, not suggestions - when teaching someone the system, start here.

RuleWhat it means in practice
Measure, don't assumeUncertain claims get flagged "this needs measurement" with a verification plan, not a confident guess.
ROI baselineEvery automation records the human baseline first: person-hours at $25/hr vs agent time + token cost, per action and at expected volume. Feeds the agent factory.
Plan before buildingPlan mode for architectural or multi-file work. If something goes sideways mid-build: stop, re-plan.
Validate smallTest connections before building on them; small-scale run before the full run.
Delegate when it winsSubagents for genuinely parallel or noisy work only. Handoffs are lossy; small-task parallelism is usually negative ROI.
Simplicity + surgical editsRoot causes over patches. Every changed line traces to the request. "Would a staff engineer approve this?"
Concise output~4-line answers, result first. The cap never applies to work products (specs, code, docs) or decision-changing facts (blockers, risks, unfinished scope).
Prove completionNever mark a task complete without showing it works. Corrections get captured to Open Brain as learnings.

Standing corrections

Durable fixes from past mistakes - the system is told these every session so it never re-learns them the hard way:

CorrectionDetail
"Sales reps"Canonical Empower term. Never "Energy Advisors" (deprecated 2026-05-15).
Granola summariesAction items are unreliable - pull the verbatim transcript before publishing any shareable meeting doc.
Brave browserDefault for scraping / CDP attach, not Chrome or Chromium.
Headless jobslaunchd claude -p jobs need --permission-mode dontAsk.
Doc reviewsMarkdown reviews go through the Obsidian vault _Reviews/ folder; edits apply to the canonical source.
Credentials rule: every API key lives in the macOS Keychain, managed by python3 ~/.keys/keys.py (get / set / list / gen-mcp / gen-codex). Nothing is ever hardcoded, and MCP configs are regenerated from Keychain, never edited by hand.

3. The dev cycle (vertical axis)

The main flow for any real build. Each step is a skill; each hands clean context to the next. Keep grill through tickets in one unbroken conversation, then give every /implement a fresh one.

On-ramps that merge into the flow instead of starting at grill:

/diagnosing-bugs - hard or intermittent bugs. Reproduce red first, then hypothesize.

/triage - externally-raised issues and PRs, moved through a 5-state machine to an agent-ready brief.

/wayfinder - work too foggy or big for one session. Maps decision tickets, resolves them one at a time, then hands off to /to-spec.

/prototype - when a design question needs a runnable answer before committing. Single HTML file, kept on a prototype/<name> branch.

When to grill: whenever the done-condition isn't checkable as stated. Small enough for one session? Skip spec/tickets and go straight to /implement. Not sure which skill fits? /ask-matt is the router.

4. Boards and modes (horizontal axis)

/so is the opt-in switch for multi-agent work. It never guesses - a cheap triage subagent scores the task and returns one of three verdicts:

Verdict 1

Pass-through

Questions, lookups, single-spot edits. Zero overhead - executes normally.

Verdict 2

Needs grilling

A real build, but an undecided decision would change what gets built. Routes to a grill interview first.

Verdict 3

Orchestrate

2+ of: parallelizable, long-running, trail-worthy, multi-specialty, boundary-spanning. Builds a Solo board: one todo per workstream, findings scratchpads, workers assigned.

House mode vs tracked mode

One file decides who owns spec-level work in a repo: docs/agents/issue-tracker.md at the project root.

House mode (file absent)Tracked mode (file present)
WhereThe normal case - all Desktop projects.Repos that ran /setup-matt-pocock-skills. Real example: JobFlo-co/jobflo (Linear tracker, since 2026-08-05).
Spec work lives inSolo todos + scratchpads.The repo's own tracker. Read the file in full before creating anything.
Project contextCLAUDE.md + .claude/history.md per the /normalize standard.Whatever the tracker doc says (jobflo: AGENTS.md is authoritative).
Solo's roleOwns the whole board.Only one agent's in-flight mechanics beneath a ticket. Never mirror tracker tickets as Solo todos.
Ownership boundary: JobFlo-co/empower-agents is Kalvin's build - we consume tools it ships, we never implement, commit, or push there. Mode is detected from the tracker file only, never from a git remote or a project's name.

5. House skills catalog

46 custom skills at ~/.claude/skills/, most symlinked from the Obsidian vault. Invoke by typing /name, or just describe the job - well-described skills trigger themselves. Click any card for what it does, when to reach for it, and a real example prompt.

Engineering workflow13 skills - running the machine itself
/soroute a task: solo board, grilling, or pass-through

WhatTriages a task via subagent, then either executes normally, sends you to a grill interview, or builds a Solo board (todos + scratchpads + parallel workers). Tracker-aware.

WhenYou explicitly want multi-agent orchestration or a visible work trail. Never fires on its own.

/so audit all six report generators for the schema change and fix each one
/wrapsession-end: commit, capture, log

WhatAuto-detects how heavy the session was, then commits code, captures learnings to Open Brain, and updates the project's .claude/history.md.

WhenEnd of any real working session. This is what keeps history.md trustworthy.

/wrap
/normalizebring a folder up to house standard

WhatAudits or fixes a project folder against the workspace standard: CLAUDE.md/AGENTS.md, .claude/history.md, naming, output dirs. Check mode (read-only report) and apply mode (approval-gated).

WhenNew project folder, or a messy old one you're about to work in seriously.

normalize this project - check mode first, show me the drift
/check-mcpshealth-check MCP servers

WhatPings every configured MCP server with a lightweight call, reports alive/dead.

WhenBefore high-stakes sessions, or when tools seem to be missing.

run check-mcps before we start the demo
/model-releaseroll a new AI model through the stack

WhatWhen any provider ships or retires a model: researches what changed, audits every local surface referencing models, applies updates in cost-of-delay order, smoke-tests, records state.

WhenNew model announcement, deprecation email, or the monthly deadline sweep flags one.

Anthropic just shipped a new Sonnet - run model-release
/skill-creatorcreate, improve, benchmark skills

WhatBuilds new skills from scratch, optimizes existing ones, runs evals, and tunes descriptions so they trigger correctly.

WhenA workflow repeats often enough to be worth packaging, or a skill keeps mis-triggering.

create a skill that builds our weekly branch report from the dashboard repo
/mcp-builderbuild high-quality MCP servers

WhatGuided construction of MCP servers (Python FastMCP or Node SDK) with well-designed tools.

WhenWrapping an external API so agents can use it - like the Scanifly and Permits connectors were built.

build an MCP server wrapping the SolarEdge monitoring API, read-only tools first
/webapp-testingPlaywright toolkit for local web apps

WhatDrives a browser against local apps: verify flows, capture screenshots, read console logs.

WhenConfirming frontend changes actually work, debugging UI behavior.

verify the login flow at command-center.test and screenshot the dashboard
/cbcommand bridge - the agent office router

WhatEntry point for the "5+3 agent office" - loads project context and routes work to the right agents.

WhenStarting structured multi-agent work inside an office-enabled project.

cb: load the empower-permits office and pick up where we left off
/design-handoffprototype to structured Linear project

WhatTurns a vibed prototype into a structured Linear project: docs, phased issues, dependencies.

WhenA prototype proved the idea and the build should become tracked, plannable work.

turn this prototype into a Linear project with phased issues
/kanbanCommand Center kanban board

WhatViews and moves tasks on the Command Center kanban, triggers Obsidian sync.

WhenManaging the personal task board without opening the app.

move "AHJ scraper" to In Progress and show me what's in Review
/linear-pmLinear project management assistant

WhatIssue creation, backlog triage, and status reports against Linear.

WhenPM chores on any Linear workspace you own.

triage my Linear backlog and draft Monday's status report
/project-kickoffstandardized project lifecycle

WhatScaffolds, tracks, and completes projects with a consistent lifecycle.

WhenStarting a net-new project that deserves proper structure from day one.

kick off a new project for the EnFin document packager
Empower & solar9 skills - the day job
/cscommercial solar project management

WhatCreates commercial project folders from intake data, evaluates go/no-go criteria, checks status, generates Jobflo reports.

WhenAny commercial solar intake, approval/denial call, or pipeline report.

cs intake for the new warehouse lead - here's the intake form data
/resiresidential solar customers

WhatManages residential customers: folder from template, branded HTML deliverables (problem page, solution page, summary page), pipeline tracking.

WhenAnything in Solar Sales/Residential - a new customer, a PPA page, an energy-cost analysis.

build the problem page for the new Elk Grove customer from her usage data
/utilityapipull authorized utility usage data

WhatFinds a customer's UtilityAPI authorization and pulls bills + interval usage as normalized CSVs into the project folder. Read-only.

WhenStarting usage or financial analysis for any customer who authorized data sharing.

pull usage for Don Anderson - bills and intervals into his folder
/dialpadcustomer call and SMS history

WhatPulls a customer's Dialpad call and SMS history by phone number. Read-only.

WhenReconstructing what was said to a customer, or checking contact history before an escalation.

what calls and texts do we have with the Morning Star contact this month?
/podiummessaging, reviews, guarded SMS send

WhatPulls Podium conversations (SMS/webchat/social), reviews, and team stats. Can send a text - but only after showing the exact message + recipient and getting explicit approval.

When"Did they text us?", review summaries, or sending a customer message through the official channel.

show me recent 1-star reviews and how we responded
/sop-architectbrain dump to standardized SOP

WhatTransforms rough notes, transcripts, or outdated docs into SIPOC-format SOPs in Empower wiki style, with versioning and role ownership.

WhenDocumenting any verbal process, or refreshing a stale SOP.

turn this meeting transcript into the install-handoff SOP
/monthly-reportbranded monthly accomplishments doc

WhatBuilds the Empower-branded monthly accomplishments Word doc from meeting notes.

WhenEnd of month, before the systems report is due.

build July's monthly report from my meeting notes
/empower-sales-chartsbranded sales visualizations

WhatChart.js visualizations in Empower brand style for sales data.

WhenAny chart going into a deck or doc that leadership will see.

chart branch sales by month for the leadership deck
/pto-design-correctionsPTO corrections worklist report

WhatGenerates the "PTO Design Corrections" HTML worklist report.

WhenThe recurring corrections review cycle.

generate this week's PTO design corrections report
Personal infrastructure13 skills - brain, comms, meetings
/brain-capturesave a thought to Open Brain

WhatCaptures thoughts/decisions/learnings to Open Brain with taxonomy-aware project routing (the whitelist in ~/.brain/taxonomy.json).

When"Capture this", "remember this", any decision worth finding again in six months.

capture this: we locked the commercial pricing model at verbal-only, no doc dollars
/sunday-reviewthe weekly brain review interview

WhatConversational weekly interview compressing the week into durable knowledge, graph links, strategy, and a one-outcome plan.

WhenSundays. Says so in the name.

Sunday review
/morning-briefingdaily unified briefing

WhatOne briefing across calendar, email, Slack, and tasks.

WhenStart of the workday.

morning briefing
/meeting-syncGranola meetings into Obsidian + brain

WhatPulls Granola meeting notes via API, writes per-meeting Obsidian files with summaries and action items, appends to the daily note, captures to Open Brain.

WhenEnd of a meeting-heavy day, or catching up after travel.

sync today's meetings
/meetingprocess one recording or transcript

WhatUnified meeting processing - audio transcription or raw text into structured Obsidian notes.

WhenA one-off recording or pasted transcript that Granola didn't catch.

process this recording into structured meeting notes
/emailGmail across both accounts

WhatGmail with strict work/personal separation (@work = Empower, @personal = gmail).

WhenReading, searching, or drafting mail without leaving the terminal.

any unread from Nikki on @work about EnFin?
/slackread, search, post

WhatSlack integration - unreads, channel history, search, posting.

WhenCatching up on channels or posting a summary.

catch me up on #empower-asks since yesterday
/calendarGoogle Calendar + meeting prep

WhatEvents, availability, and meeting preparation from Google Calendar.

WhenScheduling questions or pre-meeting context pulls.

what does Thursday look like, and prep me for the 10am
/drive-triageGoogle Drive filing

WhatTriage and file Google Drive contents - quick check or full classify.

WhenThe Drive inbox folder has piled up.

triage my Drive inbox folder, quick pass
/panning-for-goldmine a transcript for actionable threads

WhatMines a transcript or brain dump for actionable threads and captures the verdicts to Open Brain.

WhenA long rambly recording probably contains three real ideas.

pan this voice memo transcript for gold
/memory-synthesisweekly session-log synthesis

WhatSynthesizes captured session logs into structured knowledge on a weekly cadence.

WhenPart of the weekly maintenance routine.

run memory synthesis for this week
/brain-syncskills to Open Brain

WhatSyncs skill definitions into Open Brain so claude.ai/ChatGPT/Grok sessions can discover them.

WhenAfter adding or changing skills; also runs on the Sunday schedule.

brain-sync the skills manifest
/vault-maintenanceObsidian vault health

WhatVault health check and cleanup (orphans, naming, structure), quick or full, delegating to scanner subagents.

WhenMonthly, or when the vault starts feeling untrustworthy.

quick vault check
Utilities11 skills - documents, research, publishing
/docxWord documents

WhatCreate, read, and edit .docx - TOCs, headings, letterheads, tracked changes, find-and-replace.

WhenAny deliverable that must land as a Word file.

turn this outline into a formatted Word memo with our letterhead
/pdfPDF everything

WhatRead, extract, merge, split, rotate, watermark, form-fill, and OCR PDFs.

WhenAny .pdf in or out.

merge these three permit PDFs and OCR the scanned one
/gwsGoogle Workspace CLI toolkit

WhatTasks, Docs, Sheets, Drive, Chat, Meet - the whole Workspace surface from the CLI.

WhenReading or writing Workspace content programmatically.

append this week's numbers to the commissions sheet
/researchFabric AI - videos and articles

WhatRapid research via Fabric AI patterns: process YouTube videos and articles, extract wisdom, summarize, route output to the right project folder.

WhenMedia and URLs. For repo-scoped primary-source research, use the plugin's research skill instead (see section 6).

extract wisdom from this YouTube talk and file it under the agent factory
/last30days30-day multi-source research

WhatSweeps Reddit, X, web, and GitHub for the last 30 days on a topic, producing copy-paste-ready prompts and findings.

When"What's the current state of X" questions where recency matters.

last 30 days on SolarAPP+ adoption in California AHJs
/sharepublish a branded web page

WhatPublishes project content as a branded, deployed web page - ephemeral or permanent.

WhenA customer or teammate needs a link, not a file. (Proposals like Lincoln Center ship this way.)

share this proposal as a password-gated page, 30-day expiry
/export-pipelineObsidian notes to branded output

WhatExports Obsidian notes to PDF, DOCX, or email HTML with Empower branding.

WhenA vault note needs to leave the vault looking professional.

export this note as a branded PDF for the ops team
/excalidrawdiagrams in the vault

WhatExcalidraw diagrams as .excalidraw.md files in Obsidian, using Empower brand colors.

WhenA flow or architecture needs a picture that stays editable.

diagram the funding milestone flow
/google-docs-formattingprofessional Google Docs

WhatFormatting standards for professional Google Docs.

WhenA shared gdoc has to look like an adult made it.

format this doc to our standard before I share it with leadership
/nlmNotebookLM transformations

WhatNotebookLM-powered content transformation: podcasts, slide decks, infographics from source docs.

WhenRepackaging dense material into a listenable or presentable form.

make a podcast episode from this training spec
/claude-apibuild with the Claude API / SDK

WhatGuidance for building apps on the Claude API, Anthropic SDKs, or the Agent SDK.

WhenWriting code that calls Claude - agents, tools, batch jobs.

add tool use with retries to this Python agent

6. Plugin skills: the mattpocock set

25 skills from the mattpocock-skills plugin (upstream: github.com/mattpocock/skills - auto-updates, currently v1.2.x). This is the vertical axis from section 3, plus its supporting cast. Philosophy: small, composable, hackable - the human keeps control; no framework owns your process.

The main flowidea to shipped, in order
/ask-mattthe router - start here when unsure

WhatMaps every skill and flow and how they connect; tells you which one fits your situation.

WhenYou know there's a right skill but can't remember which.

/ask-matt I have a half-decided feature and a failing test, where do I start?
/grill-with-docssharpening interview, writes docs

WhatRelentless interview about a plan or design. Stateful: writes findings into CONTEXT.md and ADRs as it goes, so the next agent starts smarter.

WhenThe done-condition isn't checkable yet and you're inside a repo with (or that deserves) domain docs.

/grill-with-docs I want to add crew-conflict blocking to the calendar
/grill-mesame interview, no docs

WhatThe same relentless interview as grill-with-docs, but stateless - no files written.

WhenNon-code planning, or thinking that isn't tied to a repo.

/grill-me should we move the marketing dashboard onto Kalvin's platform or wait?
/to-specconversation to spec

WhatSynthesizes the current conversation into a spec and publishes it to the repo's tracker. No new questions - pure capture.

WhenRight after a grill, while the decisions are still in context.

/to-spec
/to-ticketsspec to tracer-bullet tickets

WhatSplits a plan or spec into tickets that each cut a narrow, complete, verifiable path through every layer, with explicit blocking edges.

WhenMulti-session builds. In house mode, don't run it unless asked - it would mint a competing board.

/to-tickets
/implementbuild one ticket, disciplined

WhatImplements one ticket or spec: drives /tdd internally slice by slice, closes with /code-review before committing.

WhenA ticket is ready and you have a fresh context for it. The default worker prompt in tracked-mode orchestration.

/implement JOB-1234
/tddred-green-refactor

WhatTest-driven development, one vertical slice at a time. Usually invoked by /implement, works standalone too.

WhenBuilding features or fixes test-first.

/tdd add the overlap validator for two customer appointments
/code-reviewtwo-axis diff review

WhatReviews changes since a fixed point on two axes in parallel subagents: Standards (repo conventions) and Spec (does it match the originating issue).

WhenClosing an implement, reviewing a branch or PR.

/code-review since main
On-rampsother ways into the flow
/diagnosing-bugsdiscipline for hard bugs

WhatRefuses to theorize until it has a tight feedback loop reproducing red, then minimize, hypothesize, instrument, fix, regression-test. Redacts secrets in captured artifacts.

WhenHard, intermittent, or "impossible" bugs and performance regressions.

/diagnosing-bugs the calendar throws "events exist for this date" only on reschedules
/triageexternal issues to agent-ready briefs

WhatMoves externally-raised issues/PRs through a 5-state machine (needs-triage, needs-info, ready-for-agent, ready-for-human, wontfix), producing agent-ready briefs.

WhenBug reports and requests raised by other people, not your own tickets.

/triage the three new bug reports in the queue
/wayfindermap work too big for one session

WhatPlans a shared map of decision tickets on the tracker, resolves them one at a time (producing decisions, not code), then hands off to /to-spec once the fog clears.

WhenGreenfield or huge efforts where you can't even spec yet.

/wayfinder the payroll system rebuild
/prototypethrowaway that answers a question

WhatBuilds a single shareable HTML file (no build, no server) to answer a design question - state model feel, or several toggleable UI variants. Kept on a prototype/<name> branch.

WhenA design question needs a runnable answer before you commit to a build.

/prototype does drag-to-reschedule feel right with 15-minute snapping?
/improve-codebase-architecturefind deepening opportunities

WhatScans a codebase for module-deepening opportunities, presents an HTML report, then grills through whichever candidate you pick.

WhenCodebase health work between features, or after a painful bug hunt.

/improve-codebase-architecture scan the inbox agent repo
Support layersvocabulary, setup, utilities
/setup-matt-pocock-skillsflip a repo to tracked mode

WhatOne-time per-repo setup: picks the issue tracker, triage labels, and domain-doc layout; writes docs/agents/issue-tracker.md + triage-labels.md + domain.md. This file is exactly what /so checks for tracked mode.

WhenOne of our own repos graduates to needing a real tracker. (This is what Kalvin ran on jobflo.)

/setup-matt-pocock-skills
/domain-modelingpin down the project's language

WhatBuilds and sharpens the project glossary - challenges fuzzy terms, records ADRs, keeps CONTEXT.md current.

WhenTerminology keeps drifting, or before a big design conversation.

/domain-modeling pin down what "milestone" vs "status" means in this repo
/codebase-designdeep-module vocabulary

WhatShared vocabulary for designing deep modules: module, interface, depth, seam, adapter. Other skills speak it.

WhenDesigning a module interface, deciding where a seam goes, making code testable.

where should the seam go between the Jobber sync and the notifier?
researchrepo-scoped primary-source research

WhatDelegates reading legwork to a background agent against primary sources; leaves a cited markdown file in the repo. Feeds grilling with grounded facts.

WhenDocs or API facts need gathering for a build. Note the name collision: house /research is Fabric AI for media/URLs. Pick by job, not name.

research how Linear's API models sub-issue blocking, write it to docs/research/
/resolving-merge-conflictsmerge conflicts by intent

WhatWorks an in-progress merge/rebase hunk by hunk, tracing each side's intent to its source. Never --abort.

WhenA rebase left you in conflict hell.

resolve this rebase - main moved the validator while my branch renamed it
/wizardbash wizard for human-only steps

WhatGenerates an interactive bash wizard walking a human through steps only they can do: provisioning, credentials, dashboard clicks, one-off migrations. Opens URLs, captures values, writes .env / secrets.

WhenThe build hits a step that needs your hands, and you want it repeatable for the next person.

/wizard set up the Cloudflare token and GitHub secrets for this worker
grillingthe interview engine (model-invoked)

WhatThe reusable interview primitive under grill-me, grill-with-docs, triage, and wayfinder. You rarely call it directly.

WhenAutomatically, whenever a skill needs to grill you.

(invoked by other skills)
writing-for-agentsreference for writing skills and CLAUDE.md

WhatReference for writing documents agents consume: skills, AGENTS.md, CLAUDE.md. Model-invoked when editing those files.

WhenCreating or editing any skill or agent-context file.

(auto-loads when editing a SKILL.md)
Productivitynon-code workflow
/handoffportable context for the next session

WhatCompacts the conversation into a portable markdown handoff doc for a new session, directory, or colleague.

WhenNear the context limit mid-phase - hand off to a fresh session rather than /compact.

/handoff we're mid-spec, next session finishes tickets
/teachmulti-session learning workspace

WhatA teaching workspace that persists across sessions while you learn a concept.

WhenYou want to actually learn something, not just have it done for you.

/teach me how pgvector similarity search actually works
/to-questionnairemine someone else for decisions

WhatThe inverse of grill-me: interviews you about the send (who it's for, what you need back) and produces a questionnaire for someone else to fill in.

WhenDecisions are blocked on another person's answers - like the definitional questions owed by a stakeholder.

/to-questionnaire Gene needs to define "Designs Completed" and the backlog buckets
/wait-whatone-word corrective

WhatFire it the instant a response doesn't land - the agent re-pitches the same content using the project's own vocabulary.

WhenYou read a reply twice and it still didn't click.

/wait-what
Other installed plugins (used automatically, no learning required): feature-dev (explorer/architect/reviewer agents), code-review, hookify (turn "never do X again" into enforced hooks), context7 (live library docs), playwright, security-guidance, frontend-design, laravel-boost, github, commit-commands, and the private empower-funding plugin (EnFin packages).

7. Memory and brain

Two stores, different jobs, both feeding each other:

File memory

MEMORY.md + topic files

Always-loaded index with a rotating "Active focus" list, plus 8 topic folders (profile, feedback, setup, work_empower, projects, reference, ventures, retired), each with its own on-demand index. One markdown file per fact. Fixed token cost no matter how much history piles up.

Open Brain

Semantic search over everything

Supabase + pgvector store queried with brain_search before answering any "did we already decide..." question. Captures typed as thought / learning / decision / project / reference, tagged against a project taxonomy whitelist.

The loop: work happens, /wrap logs it to the project's history.md and captures learnings, /sunday-review compresses the week, /memory-synthesis structures the logs, and corrections become standing rules so mistakes don't repeat. When teaching someone: the habit that matters most is capturing decisions when they happen - everything downstream depends on it.

8. Models and research routing

Model stratification

Use the simplest model that can do the job. This is the biggest cost lever in the whole system.

ModelRole
Fable 5Top tier (Mythos-class). Hardest reasoning, highest-stakes plans, boss/checker/verdict roles. Not the everyday default.
Opus 5The daily driver: architecture, complex multi-file changes, nuanced writing, most real work.
Sonnet 5Standard subagent work: research, code review, routine development, most delegated tasks.
Haiku 4.5Mechanical tasks: formatting, simple lookups, file operations.

Research routing

NeedTool
Quick factual lookup, reasoningGrok CLI headless, first choice
Deep multi-source researchGrok, then WebSearch
Fetch a known pageFirecrawl or WebFetch
Video or article processing/research (Fabric AI)
Library and framework docscontext7 plugin
Past decisions and historybrain_search - always before re-deciding anything

Key paths

PathWhat lives there
~/CLAUDE.mdThe global contract - everything in section 2.
~/.claude/skills/All 46 house skills (section 5).
~/.keys/Credential management via keys.py; Keychain-backed.
~/agent-stack/Playbooks (cross-portal browser automations) + portals config.
~/Herd/Laravel projects; ~/Herd/CLAUDE.md auto-loads stack conventions.
~/Desktop/<project>/Project workspaces: CLAUDE.md + .claude/history.md each.