Adam Hasham

Founder, ex-Amazon/Alexa. Building agent-driven personal knowledge graphs in public. Vault patterns, skill specs, single-writer zones.

[Adam Hasham @ahasham]

My personal knowledge graph: an INPUT layer (Gmail / calendar / WhatsApp) and an OUTPUT layer (Career Pipeline + Family Todo). Claude skills are the transforms.

A dbt-style medallion architecture for personal data.

Here's why this is the only PKM design that survives AI agents.

[Adam Hasham @ahasham]

You know you’re getting close to the token limit for the week when...

[Adam Hasham @ahasham]

@JqOnly @sidewalkco Same shape in my personal-OS. Stage 7 has 4 deterministic gates before publish: text matches, button enabled, share-safety clean, counter > 0. OS defines the Handoff Boundary, LLM produces inside it. Without those contracts, the skill works 80% of the time.

[Adam Hasham @ahasham]

@Wattenberger This generalizes past code. My personal-OS has the same shape: one skill per vault zone (Career, Family, Health), each with bespoke editor + bespoke substrate. The skill that owns the zone makes the substrate worth editing. What does this look like in your sandbox?

[Adam Hasham @ahasham]

The 'specs in YAML' framing on HN reverses what works in my Claude skills. The spec is prose the model executes; YAML is identity + tunables (name, paths, defaults). Conflate them and you lose prose's expressivity AND config's stability.

[Adam Hasham @ahasham]

git has tracked Co-Authored-By trailers for years. VS Code finally autopopulating them with the LLM author is obvious in retrospect. Every Stage M auto-commit in my personal-OS lands with `Co-Authored-By: Claude opus-4-7[1m]`. What else are we leaving on the floor?

[Adam Hasham @ahasham]

@ApplyWiseAi Fragmentation isn't recall. Vectors solve that. It's a write problem. Single-writer per zone: exactly one skill writes Career, exactly one Family Todo. Cross-zone updates are explicit handoffs in config. "Two agents fight over one note" goes away.

[Adam Hasham @ahasham]

AWS shipped Amazon Quick. Big-tech just validated the personal-knowledge-graph category. Here’s the bet that stays mine.

[Adam Hasham @ahasham]

@shreyasmakes Did this. 25 skills now. Three surprises: (1) the chassis matters more than any single skill, uniform Stage M is what makes self-learning compound; (2) single-writer per file beats per-directory for personal data; (3) hooks beat memorization for system contracts.

[Adam Hasham @ahasham]

@web3nomad totally agree! and the output layer is both - they're md files (I use @obsdmd) so I can click checkboxes and tag @claude to do things but I also have apps on top that read those files and add a UI layer, features, etc.

[Adam Hasham @ahasham]

My personal knowledge graph: an INPUT layer (Gmail / calendar / WhatsApp) and an OUTPUT layer (Career Pipeline + Family Todo). Claude skills are the transforms.

A dbt-style medallion architecture for personal data.

Here's why this is the only PKM design that survives AI agents.

[Adam Hasham @ahasham]

Open Q for anyone building this kind of system: when should an agent re-surface something I deleted weeks ago?

My current heuristic is "never unless an input explicitly contradicts the prior delete." Probably wrong in the long tail.

[Adam Hasham @ahasham]

Closest patterns: Karpathy's LLM Wiki (forbids human writes), Steph Ango's obsidian-skills (sacred areas).

The wedge: edits-as-input + zone-ownership + multi-skill specialization.

[Adam Hasham @ahasham]

The fix: don't lock me out, don't merge.

I edit the output freely. The agent regenerates everything from scratch, reading my prior output (with my edits) as input.

A checked box is signal. An @claude marker is an instruction. A deletion stays deleted.

[Adam Hasham @ahasham]

Most "AI second brain" tools fall into one of two failure modes:
(a) lock you out of the agent's output, or
(b) try to merge with your edits and corrupt them.

You can see this play out today in Cursor, Kiro, and even VS Code's strict-mode debate.

[Tarun Sachdeva @tarunsachdeva]

Introducing Share from Anywhere for Traces

A set of skills for sharing to directly from any supported agent to @tracesdotcom. Available for all 9 supported agents.

This small add-on has improved how we work with agents as a team. Here's how it works: