Correction loop
Public claims got tighter
A March article opener was narrowed into a defensible 2026 claim, then follow-up teasers changed to match.
memory/compound-reviews/2026-03-26.mdHiggins
Designing an AI collaborator I could live with. A scheduled, memory-backed, permission-bounded system that turns saved context into briefs, drafts, tasks, and repair loops.
RoleDesigner + Builder
ToolsClaude Agent SDK, Claude Code
Year2026
ImpactSaved signal becomes decisions
Coordinator
Higgins
Routes work to specialists and reads from the shared memory layer.
Specialists
TedSherlockLesterMikeGilfoyleRonTonyMs_MaiselDavidScruffy
Memory layer
Active memoryObsidianTodoistSweep jobs
292 memory files.22 scheduled tasks.11 specialists.Claude Agent SDK.running since Jan 2026
Context
Saved signal kept becoming another backlog
Twitter saves, bookmarks, client threads, project notes, invoices, and half-remembered commitments were scattered across too many places.
Hard Part
Trust was the product work
Higgins had to know what he could touch, what needed review, and when a scheduled surface was welcome.
Breakthrough
Short memory for today. Obsidian for the rest.
Active memory stays slim. Durable lessons get swept into Obsidian so the next brief starts with more context.
What Shipped
A set of surfaces I keep using
Morning briefs, bookmark digests, public essays, repair loops, Todoist cleanup, portfolio planning, and small business tools.
Signal Pile
I had more signal than I could turn into decisions.
I save too much. Tweets, articles, product ideas, little patterns from client work, things I should follow up on later. Left alone, it turns into a guilt pile.
Higgins started getting useful when that pile became an operating surface. Saved links could become a brief. Loose thoughts could become drafts or tasks. Broken infrastructure could become a repair note instead of a thing I had to remember at midnight.
Morning Brief
Higgins turns the stuff I save into advice before the day starts.
The Morning Brief pulls from weather, Todoist, bookmarks, scout signals, self-heal status, memory, and active project context. A nicer summary is table stakes. I wake up to useful context I didn't have to reassemble.
Source receipt, May 25
Weather, Todoist, saved signal, repair status, and project context in one daily briefing.
This is why the podcasts matter. Higgins can turn saved posts and bookmarks into a brief that says what changed, why it maps to my work, and what I might do differently.
Voice
Personality lowered the cost of context. Correction made it compound.
The character matters because I return to him. It's easier to give context, easier to push back, and easier to let the system remember what changed. That's how correction becomes future behavior instead of another old transcript.
Memory
The breakthrough was memory that could get cleaned up.
Short-term memory keeps today's work close. Obsidian keeps the durable stuff readable. Sweeps move useful context between them so advice gets better without dragging every old note into every turn.
Memory loop
Useful signal survives the bookmark.
01
Saved signal
Bookmarks and scout notes catch language I would not have time to unpack one by one.
02
Digest
Higgins summarizes the source and mirrors the useful part into memory.
03
Long memory
Obsidian keeps durable knowledge browseable, editable, and separate from task truth.
04
Playbook
The idea becomes reusable guidance, like transition vocabulary for future UI work.
05
Morning Brief
The same signal resurfaces as advice when it can change what I do that day.
Source: memory/briefings/bookmark-digest-2026-05-25.md, 400 Resources/Playbooks/animation.md, 100 Journal/Daily/2026-05-25.md
Trust
Architecture matters because each layer protects the relationship.
This is where the product decisions show up. Active memory keeps today's work close. Obsidian keeps durable knowledge browseable. Todoist holds action. Email stays draft-only until I say yes.
Memory
Active files
Obsidian vault
Sweep jobs
Trust
Todoist task truth
Draft-only email
Scoped tools
Repair
Self-heal ledger
Session reset
Continuity note
Trust architecture
Boundaries made delegation usable.
Todoist
Tasks live in one place
If work needs follow-through, it leaves notes and lands in the task system.
Obsidian
Knowledge stays browseable
Research, digests, playbooks, and project memory stay readable outside the agent runtime.
Communication stays draft-only
Higgins can prepare language. Sending still waits for an explicit yes from me.
Permissions
Tools are scoped by job
Specialists get the access their lane needs, which keeps delegation boring enough to trust.
Source: agents/default/references/task-management.md, agents/default/references/obsidian-sync.md, agents/default/AGENTS.md
Surfaces
The proof is the range of work Higgins touches.
Briefings are the cleanest proof, but the mundane stuff matters too. Higgins has helped with proposal strategy, invoice generation, research synthesis, Todoist migration, writing drafts, and this portfolio rewrite.
Admin
Proposal and invoice surface
Proposal work moved from brittle DOM capture to generator-backed preview/export, then a final PDF got visually inspected.
memory/handoff-log.md, 2026-03-20Strategy
Portfolio proof selection
Higgins helped decide what belongs on this page, then pushed the story toward receipts instead of vibes.
memory/2026-05-25-1211.mdWriting
Research became drafts
Article research, community evaluation, and draft artifacts were produced and mirrored to Obsidian.
memory/compound-reviews/2026-03-20.mdRepair
Repair loops are where the system stops being a pile of automations.
The safer claim is narrower: Higgins detects common failure modes, performs targeted resets when evidence is present, and logs blockers when repair would be unsafe.
Repair loop
Failures become named repair events.
Detect
Drift becomes visible
A session-model mismatch or timeout lands as evidence with a reason, timestamp, and follow-up note.
Repair
Reset the surface
The helper resets the targeted session and writes continuity instead of restarting everything.
Learn
Write the rule
The failure class gets named so the next sweep can catch it earlier.
Source: memory/2026-05-17.md, memory/2026-05-20.md, memory/errors.md
Outward
The same loop works when the stakes move outside my notes.
External collaboration is the privacy-sensitive proof, so the page only needs the shape: interpret, draft, review, learn, send after approval. This case study is the smaller proof sitting in public.
External loop
The same boundaries extend outside my workspace.
Higgins can help interpret an incoming request, prepare a response, and turn my corrections into rules. The send step stays gated.
1. Inbound
2. Interpret
3. Draft
4. Kyle review
5. Learn
6. Approved send
Counterparty: [redacted]
Project details: [redacted]
Approval: required before send
Source: Obsidian/300 Projects/Luck Agents/Communication Log.md and Communication System.md, redacted for publication
Recursive proof
The case study became part of the loop.
Critique
Higgins pushed the draft away from agent inventory and toward product judgment.
Evidence
He gathered receipts before the tone pass, then flagged which claims were risky.
Revision
The page now uses Higgins as proof of the working relationship it describes.
Source: memory/2026-05-25-1211.md, memory/2026-05-25-1332.md, memory/2026-05-25-1435.md
Outcome
Less re-explaining. More operating with continuity.
Saved links become decisions. Loose thoughts become briefs, drafts, and tasks. Failures become rules or repairs. Agent count is metadata. The useful part is that I spend less energy rebuilding context and more energy deciding what matters next.
The weirdest proof is this page. Higgins critiqued the Higgins case study, found the weak claims, and pushed the story toward receipts. That's the working relationship I was trying to build.