agentleFS
NVIDIAKnown publisherNVIDIA/nemoclaw-community/examples/recipes/nvidia/memory-driven-chief-of-staff/profile/skills/inbound-judging/SKILL.md

inbound-judging

Skill142 starsChanged 3 months agoApache-2.0View on GitHub

Decide which newly-arrived email and Slack messages create an obligation for the user, and craft the row that represents each one.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    A stale current_priorities.md still counts: If the page is past its decay window, use it anyway and note in urgency_reason that the …

    line 119
  2. Very specific:

    Every message in the slice appears exactly once.

    line 147

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Judging inbound messages

You receive a slice of messages that have never been judged, plus the user's currently open obligations. Decide which messages create an obligation, craft a row for each, and return one JSON envelope. You never write to the database — scripts/apply_decisions.py does that.

Read first

  • $HERMES_HOME/workspace/memory/index.md, then only the pages it names that bear on this slice — typically people/ for the senders and the relevant projects/.
  • $HERMES_HOME/workspace/policy/preferences.md if it exists. It records what this user has repeatedly ignored or re-ranked. Treat it as a strong prior, not a rule: a pattern ignored eight times is probably noise, but an unmistakably urgent instance of it is still an obligation.

Stage 1 — flag

Mark a message when any of these hold:

  1. The sender's people/<slug>.md has importance: high.
  2. addressing is direct or mentioned, and the body carries a priority signal — a deadline, a date, "before you leave today", an explicit ask. A broadcast item needs a stronger reason than urgency alone; being sent to a list is not being asked.
  3. unread is 1 and the sender is a person rather than a noreply@ / do-not-reply@ style address. Slack items leave unread empty, so this rule simply does not apply to them.
  4. It refers to something already represented in the open obligations — the same ticket, thread, or deliverable.

Suppress your own output. Summaries, briefings and digests this agent generated are derived text, not evidence. If one mentions a real obligation, act only when the original message is in this slice or already has a row.

Suppress the already-handled. Compare each candidate against the recently completed and ignored obligations in your input. A reminder, a re-send, an "overdue" nudge or a thank-you about work the user already finished gets no row, even though it arrived as a new message with a new id. Match on the underlying real-world obligation — same person, same artifact, same request — not on wording. The exception is a genuinely new and distinct ask: propose that new action alone, and say in context that the earlier related task was already resolved.

Everything you do not flag is SKIP. That is a terminal state, so an item is never re-judged, and the cost of a slice stays bounded. Be willing to use it.

Stage 2 — craft the row

Title

The title is the only field the user sees without opening anything. It must stand alone.

  1. At most 80 characters. Hard cap. Trim tail tags; the priority field already carries urgency.
  2. Lead with the verb the user will execute — Draft, Reply to, Review, Investigate, Decide, Schedule, Renew, Verify, Prepare, Ship.
  3. Self-sufficient. Someone reading only the title, with no source and no context, must know what to do.
  4. Never open with Work on, Coordinate with, Monitor, Track, Follow up on, Handle, Look into, Check on. Those are labels pretending to be actions.
  5. Never reuse a subject line. Translate the intent into an action.
SubjectWrongRight
Q3 capacity planningWork on Q3 capacity planningSend Q3 headcount numbers to Dana by Thursday
Re: vendor contractFollow up on vendor contractDecide whether to accept the vendor's Sept 2 window
BILL-441: migration dry runTrack BILL-441Schedule the billing migration dry run (BILL-441)

Context

At most three bullets. What the user needs in order to act without opening the source. Not a summary of the message — the parts that change what they would do.

kind

response when the deliverable is the reply itself. action when the work happens elsewhere and a reply is incidental. A message that asks for both takes the one holding the bulk of the effort; mention the other in context. There is deliberately no third value.

est_effort

minutes (5-30), hours (1-4), day (4-8), multi_day. One focused sitting, not elapsed calendar time. response rows are almost always minutes; use hours only for a genuinely substantial write-up.

When uncertain, choose the higher tier. Over-estimating costs the user nothing. Under-estimating lures them into starting something they cannot finish.

intent_gated

Set true when the obligation maps to something the user has chosen to work on, judged against the memory:

  • an entry in attention/current_priorities.md, or
  • an active project or goal page — a project is active when its updated is within 30 days and Health is not a closing state.

Set false when the message is merely urgent — a tight deadline, an important sender, a broadcast mention. External pressure is not the same as intent, and only intent earns the top tier.

A stale current_priorities.md still counts. If the page is past its decay window, use it anyway and note in urgency_reason that the priorities it reflects were last confirmed on its updated date. Out of date is not the same as wrong, and switching the gate off entirely would push every row into the middle tier — worse than acting on slightly old intent.

Stage 3 — rank and emit

Order the rows you are keeping by how much they deserve the user's attention. You decide the order; the caps and tiers are applied afterwards in code, so do not assign priorities yourself.

Emit exactly one envelope, nothing else:

{
  "version": 1,
  "pass": "intake",
  "decisions": [
    {"source_id": "...", "decision": "CREATE", "rank": 1, "intent_gated": true,
     "title": "...", "context": "...", "urgency_reason": "...",
     "kind": "response", "est_effort": "minutes"},
    {"source_id": "...", "decision": "SKIP"}
  ],
  "cursor": {"source": "email", "scope": "inbox", "value": "..."}
}

Every message in the slice appears exactly once. A message you neither create nor skip is a bug — the item stays pending and will be re-judged forever.

Then write it

The envelope is not the deliverable; the stored rows are. Save it and run the writer:

python3 "$HERMES_HOME/scripts/apply_decisions.py" < envelope.json

It prints how many rows it created, updated, closed, and skipped. Report those counts. A run that printed an envelope and never invoked the writer changed nothing, however good the judgment in it was.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.