Low-effort drop zone for anything the team thinks might be useful but doesn’t yet warrant a full research note. Dumps live here until an agent (or a human) processes them into a proper research/ note, an ADR in decisions/, or an entry in the memory layer.
What goes here
- Links to products / articles / repos worth investigating (e.g.
2026-04-20-hyperspell.md). - Patterns or ideas we’ve noticed in the wild.
- Half-formed thoughts that need research before they’re useful.
- Transcripts, screenshots, or raw snippets to be summarised later.
- Candidate tools, vendors, or libraries surfaced by the team.
Threshold is low by design. A 10-line note naming the thing + why it’s interesting is better than no note at all.
What does NOT go here
- Decisions the team has made — those go in
decisions/as ADRs. - Completed research — that goes in
research/. - Session transcripts — those go in
status/sessions/per the shared project memory note. - Secrets, credentials, or NDA-covered material. Same rules as the rest of the repo — see
CLAUDE.md.
How dumps get processed
A scheduled agent (future — candidate platforms in the Multica + Managed Agents research) picks up unprocessed dumps on a cadence and:
- Reads the dump file.
- Runs a web-research pass against the listed URLs + related searches.
- Emits a full research note into
research/YYYY-MM-DD-<slug>.mdand an HTML companion, following_template-research-output.mdfor the quality bar. - Commits directly to
mainand pushes. Research is zero-risk content — the PR ceremony doesn’t buy anything the team couldn’t recover from with a revert commit. The agent also updates cross-links (See also:back-references) in the same commit. - Flags the dump as processed by appending
**Status:** processed → [link to new note]at the top of the dump file, in the same commit.
Direct-push guardrails (even without a PR)
The agent can push directly, but it must:
- Stay in its lane. Writes permitted in
raw-dumps/,research/,status/daily/, andstatus/sessions/only. Any commit touchingdecisions/,docs/,app/,CLAUDE.md,BRANDING.md,TEAM.md,templates/,.github/, orpackage.jsongoes through a PR — not a direct push. - Use a
auto-research:commit prefix. So humans cangit log --grep='auto-research:'and spot-check or revert. - Co-author the commit to its own handle. E.g.
Co-Authored-By: auto-research-agent <noreply@oxflow.design.3sixtyone.co>. - Fail safe. If the research can’t cite the minimum sources in
_template-research-output.md, the agent commits a stub toresearch/withstatus: draft — insufficient sourcesand leaves the dump unprocessed.
Until the agent is wired in, dumps stay here and can be picked up manually (same flow; the human commits directly to main too — no PR).
Filename convention
raw-dumps/YYYY-MM-DD-<short-slug>.md
- ISO date of when it was dumped.
- Short hyphenated slug — product name, topic, or URL stem.
- Underscore-prefixed files (e.g.
_template-research-output.md) are meta, not dumps.
Minimal dump format
# <Product or topic>
**Dumped:** YYYY-MM-DD by Real Name (@handle)
**Source:** one or two URLs
**Category:** product | pattern | article | repo | transcript | other
**Status:** unprocessed | in-progress | processed → link to research note
## Why this is interesting
One or two sentences.
## Raw notes
Anything you thought while dropping this. Quotes, selections, screenshots-as-references. No structure required.
## Proposed research angle
What an auto-researcher should actually chase if this gets picked up.
## Related in-repo docs
- [Relevant research note](../research/...)
- [Relevant ADR](../decisions/...)Keep it short. The point is that the dump is the raw — the synthesis is what the agent (or the next human researcher) produces downstream.
Why this folder exists
Right now good ideas get lost in Slack DMs, read-it-later tabs, and voice memos. This folder is the committed-in-git version of that inbox — diffable, searchable, visible to the team, and ready for an agent workflow to process.
See the shared project memory research for the broader context-capture pattern this fits into.