Knowledge and workflow design

The recurring problem

I produce useful material while doing other things: talking with AI, reading research, reflecting on experience and preparing work.

A conversation can run to hundreds of pages. A paper may be worth remembering without requiring a full analysis today. An observation may acquire meaning only when seen alongside later observations. Evidence from past work can remain scattered across old proposals, reports and CVs until a new opportunity makes it relevant again.

Many knowledge-management methods add a second task to the original activity. Capture the useful item. Classify it. Decide where it belongs. Maintain the categories. Tend the garden.

That administrative interruption changes the activity and creates enough friction for useful material to disappear.

For years, the systems that promised to extend my thinking required more executive work than I could sustainably give them. AI is finally making it possible to externalise, develop, preserve and rediscover ideas without slowing my thinking to the speed of manual knowledge management.

The design approach

These systems follow the same broad approach:

• capture at a natural pause or retrospectively;
• automate repetitive classification where useful;
• retain access to the original source;
• distinguish records from later interpretations;
• support browsing and detailed retrieval separately;
• allow material to remain unfinished;
• create boundaries that suit the subject, audience and level of privacy.

Across these five systems, I have turned recurring losses of knowledge into working capture, classification and retrieval workflows. The subjects vary, but the design work is transferable: diagnosing where information disappears, reducing the effort required to preserve it, deciding which tasks to automate and creating structures that remain usable as the material grows.

The systems

System	Material	Workflow problem	Designed progression
Evidence Garden	Evidence from professional work	Useful evidence is scattered across old CVs, proposals, reports and memories	Episode → checked evidence → reusable career pattern
InfoMagpie’s Nest	Ideas recovered from long conversations with AI	Useful thinking disappears inside conversations that may run to hundreds of pages	Conversation → Pebble → connection, question or possible future work
Inner Magpie	Personal observations, recurring patterns and developing self-understanding	Reflection needs continuity, privacy and very low-friction capture	Observation → connection → accumulating personal understanding
BusyDoggie Fieldnotes	Canine research, observations and developing interpretations	Ideas often develop long before I have time to turn them into finished articles	Source or observation → fieldnote → possible future publication
Agentic Systems Reading Room	Research on agentic and intelligent systems	The volume of new research creates overwhelm, FOMO and repeated rediscovery	Discovery → triage → retrievable research record

Evidence Garden

Evidence Garden recovers professional evidence from a nonlinear career.

The system records episodes, establishes what happened and preserves the source of each claim. Patterns and recurring capabilities can then be examined across the records.

This order keeps professional claims traceable to evidence while allowing patterns to emerge across projects.

Explore Evidence Garden →

InfoMagpie’s Nest

I use ChatGPT as a thinking surface, and useful ideas often develop across very long conversations. Exports, summaries and synthesis documents helped to some extent, but material still disappeared.

InfoMagpie’s Nest provides a lower-friction route from conversation to a small, retrievable record. Automated tagging allows detailed classification without requiring me to stop and organise every record manually. The system preserves detailed metadata without transferring the maintenance burden back to the user.

A Pebble can connect with other ideas, contribute to future work or remain a Pebble.

Explore InfoMagpie’s Nest →

Inner Magpie

A private system for preserving self-knowledge across time. Selected observations, dated context and developing interpretations can be revisited without forcing them into a fixed personal profile.

Read about Inner Magpie →

BusyDoggie Fieldnotes

Ideas about canine behaviour and research often develop much faster than I can turn them into finished articles. Writing takes time, the dog blog is hardly my day job, and an interpretation may continue changing while it waits.

BusyDoggie Fieldnotes gives those ideas an intermediate form. They can remain connected to their sources, develop over time and become posts when they are ready.

Explore BusyDoggie Fieldnotes →

Agentic Systems Reading Room

I encounter new research on agentic systems through scheduled weekly research sweeps, social media and other sources. Reading every paper in detail would be a poor use of time. I still want to remain oriented, gradually develop my understanding and find relevant work again later.

The Reading Room supports triage. Each sweep adds papers and initial assessments to a cumulative research record. Each sweep makes the next one more useful because earlier papers, assessments and themes remain available for comparison.

Explore the Agentic Systems Reading Room →

How AI fits into the workflow

AI performs different functions across the workflows: extraction, classification, initial synthesis, retrieval and implementation.

This reduces the manual work involved in maintaining detailed organisation. InfoMagpie’s Nest can use granular tags without requiring manual classification of every Pebble. The Reading Room can absorb a continuing stream of research before each paper receives detailed attention.

Classification remains interpretive. The useful vocabulary changes as a collection grows, so AI-generated tags remain editable. The original material remains available because summaries and labels provide only partial representations of a source.

Human judgement determines:

• what deserves capture;
• which distinctions and sources need preserving;
• how much structure is useful;
• what belongs together;
• where a separate boundary is needed;
• whether an apparent connection is meaningful.

AI handles much of the repetitive labour. I design the system and judge its outputs.

Shared design decisions

Design challenge	Design response	Evidence from the systems
Capture competes with the activity producing the material	Fit capture around the activity already happening	InfoMagpie captures after or during natural pauses in a conversation; the Reading Room captures papers during research sweeps
Detailed classification improves retrieval but creates maintenance	Automate repetitive tagging while keeping classifications editable	InfoMagpie uses automated tags that can be corrected as its vocabulary develops
Granular metadata makes poor primary navigation	Separate browsing from detailed retrieval	InfoMagpie replaced 568 visible labels with nine broad, overlapping browsing areas while retaining granular tags for search and connection
A map of connections can obscure the material being connected	Put the records before the visualisation	An early graph-oriented InfoMagpie homepage was replaced with search, broad browsing areas and actual note titles
A tidy record can give an interpretation more authority than its source supports	Preserve sources and interpretations separately	Evidence Garden keeps claims linked to their evidence; BusyDoggie records retain the source that an interpretation questions or develops
Some records need time before their destination becomes clear	Support development without demanding an output	InfoMagpie Pebbles can remain Pebbles; BusyDoggie fieldnotes can develop into articles when enough thinking has accumulated
Different material requires different privacy boundaries	Use related methods across separate collections	InfoMagpie and BusyDoggie are public; Evidence Garden and Inner Magpie are private
A growing collection changes its own organisational needs	Allow categories and metadata to evolve	Weekly research sweeps continue adding material to the Reading Room; InfoMagpie’s browsing structure changed after the original label system became unwieldy

What this work demonstrates

Capability	Evidence in this work
Behavioural diagnosis	I identified interruption, maintenance and repeated classification decisions as the main barriers to capture
Workflow design	Capture happens during natural pauses or retrospectively, with repetitive organisation moved away from the moment of thought
Information architecture	InfoMagpie separates nine broad browsing areas from 568 granular retrieval labels
Automation judgement	AI proposes tags and extracts records, while classifications remain editable and linked to source material
Provenance design	Evidence Garden and BusyDoggie preserve the distinction between source records and subsequent interpretations
Iterative prototyping	The InfoMagpie graph homepage was removed when use showed that it obscured the material
Privacy and boundary design	Related methods support public, professional and personal collections without combining their contents

Different systems need different boundaries

One repository could create the wrong boundaries between professional evidence, personal reflection, public notes, canine research and emerging technical literature.

These materials have different audiences, sensitivities and possible destinations. They also require different kinds of context.

Inner Magpie and Evidence Garden are private. InfoMagpie and BusyDoggie are public. The Reading Room is publicly accessible but primarily serves as my own cumulative research record.

The shared method works across separate systems while each collection retains appropriate rules for privacy, evidence, interpretation and publication.

Knowledge infrastructure for co-intelligence

These projects are part of my wider work on designing co-intelligence.

A conversation can be intelligent in the moment and still lose the context that would improve the next conversation. Knowledge infrastructure gives collaboration continuity. It preserves the source behind an idea, the observations preceding an interpretation, the evidence supporting a professional claim and the questions still under development.

Inner Magpie makes this connection particularly visible. Designing support that extends someone’s intelligence requires some understanding of how that person thinks, what they notice, where they struggle and which patterns recur.

A private, accumulating record can support that understanding while leaving room for change, contradiction and incomplete interpretation.

The same principle applies to organisational AI systems. Their usefulness depends on the quality, provenance and retrievability of the context available to them, as well as the interaction taking place now.

Co-intelligence develops through conversation and through the systems that preserve, retrieve and return knowledge to its context.