Usage

1. Everyday work: nothing to type
2. ulw: one change, proven
3. mass ulw: a job split across agents
4. Knowledge-graph commands

Everyday work: nothing to type

Hindsight runs on hooks:

• On your first prompt, it injects a short synthesis of what the repo remembers.
• On every prompt, it refreshes the list of knowledge pages.
• When a session ends, it writes the session back.

To store something deliberately, say so: “store in hindsight that we round prices half-up”.

Context+ tools are available whenever the agent needs structure or meaning-based search. Exact-text search still uses Grep.

ulw: one change, proven

ulw the cart badge shows 0 after adding an item on mobile

The ultrawork skill makes the agent:

1. Find the owning code with semantic search and read it, checking memory for earlier decisions.
2. Reproduce the bug, or state the acceptance check for a feature, before changing anything.
3. Run get_blast_radius on every symbol it changes, and keep the diff to the request.
4. Prove the result with the repo’s own test, type and lint commands, run in the session, retrying the same failure at most three times.
5. Report what isn’t done first, then the evidence.

If you have your own skill named implement-change, ultrawork defers to it.

mass ulw: a job split across agents

mass ulw migrate every package under packages/ from the v1 to the v2 logger API

The ultrawork-graph skill:

1. Scouts first and applies the stop rule. If the work is sequential, or has fewer than three independent pieces, it runs ultrawork instead and tells you why.
2. Draws the graph. Each node gets a goal, the files it owns and its done-evidence. An edge exists only where one node’s output feeds another. The graph appears in the task list.
3. Runs the nodes in parallel:
   • scout (Haiku, read-only) for discovery
   • worker (Sonnet, its own git worktree) for code changes
4. Verifies each node with a separate verifier that reruns the evidence itself. A failing node is retried at most twice.
5. Merges once, then runs the repo’s checks across the merged result.
6. Asks you only before steps that are expensive to undo, such as migrations, deletions or pushes.

Knowledge-graph commands

Run them in order for a full build, since each one consumes the last one’s output:

Command	Does
/devstack:kg-tutor	Interactive course, one module per message, using your project as the example
/devstack:kg-scope	Entity and relation types; your questions rewritten as traversals
/devstack:kg-schema	A real ontology plus Turtle for Protégé
/devstack:kg-extract	Extraction pipeline per source type, failure modes, hand-check protocol
/devstack:kg-relations	Typed triples with evidence spans; a distant-supervision baseline
/devstack:kg-events	Event schema, causal and temporal edges
/devstack:kg-fuse	Entity resolution with blocking, a review band and reversible merges
/devstack:kg-eval	Triple-level precision and recall, leakage checks, baselines
/devstack:kg-rag	Retrieval per question type, proven against vector search

Arguments fill the bracketed fields: /devstack:kg-scope a catalogue of open-source licences; which licences are compatible with GPL-3.0?