Agents, Memory & Skills
How teams on this project use AI coding agents - the vocabulary, the decision method, and who owns what.
Agents, memory & skills
A shared place for how teams use AI coding agents effectively - not just CTD, but the frontend team on this repo, whoever owns the scraper repos, whoever owns the CI/CD automation in pipeline, and any team that comes after. The goal is the same as everything else here: less duplicated tribal knowledge, less onboarding by osmosis.
Almost none of this content exists yet, so this section is a method for growing it team by team, not a finished structure to fill in.
Where this repo stands today
One file: AGENTS.md, one rule - a warning that this app's pinned Next.js
version has breaking changes versus a model's training data. No skills
directory, no subagents, no hooks. That is the honest starting point; nothing
below invents practices this repo doesn't have yet.
The vocabulary - four mechanisms, one job each
| Mechanism | Lives at | For | In this repo, today |
|---|---|---|---|
| Memory | AGENTS.md / CLAUDE.md, repo root | Standing context every agent session loads automatically - the guardrails you'd otherwise repeat in every prompt | One rule (above) |
| Skill | .claude/skills/<name>/SKILL.md (or your agent's equivalent) | A packaged, reusable procedure - a checklist or workflow invoked by name mid-session, not repeated from memory each time | None yet |
| Subagent | .claude/agents/<name>.md (or equivalent) | A specialized persona with its own tool access, for a recurring task worth a dedicated role - a reviewer, a researcher | None yet |
| Hook | .claude/settings.json (or equivalent) | An automated action on a session event - run a formatter after every edit, block a risky command | None yet |
The decision method
Short version:
- Something every session must always know -> memory.
- A procedure you'd explain the same way every time -> a skill.
- A distinct role worth its own context and tool scope -> a subagent.
- Something that should happen automatically, unasked -> a hook.
Most teams starting out only ever need the first two.
The team table
| Team | Page | What exists today |
|---|---|---|
| Meetings viewer (frontend) | meetings-viewer | AGENTS.md with the Next.js version warning |
| Scrapers (CTD + PDW) | scrapers | One confirmed skill: spider-review (PDW, review-side); a write-side companion is recommended, not yet adopted |
| DevOps / CI-CD | devops | Stub - owned by whoever maintains the automation in pipeline |
The one rule that keeps this honest
The docs hold the facts; the skills hold the judgment. A skill carries
procedure, heuristics, and worked examples - the things that are genuinely its
own. For anything enumerable and generated (schema fields, constants,
commands, the QA rubric) it links to the doc anchor instead of restating it.
That works because this site publishes per-page .md routes and
llms-full.txt - an agent can follow the link and read the source text
directly.
To add your team's section, follow how to add a section.
Last updated on