agentleFS
geekxxgeekxx/ai-assistant-starter-kit/.claude/skills/sweep-binder/SKILL.md

sweep-binder

Skill1 starsChanged 4 months agoMITView on GitHub

On-demand reconciliation of the FD Program Binder / per-firm tracker gsheet against what a sweep just learned, for one or all Forward Deploy firms. Produces a NEW/UPDATE/RESOLVED-CANDIDATE proposed-changes block per firm for Jeff's approval — it never writes to a gsheet (no Sheets-write tool exists today), so the approved diff ships as a markdown file for manual application. Trigger on "sweep binder," "update the tracker," "reconcile [firm] tracker," "binder sweep.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 6 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

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

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

    Default scope is all 5 firms (per _shared/fd-firms.md).

    line 37
  2. Very specific:

    Read the FD Program Binder (mcp__claude_ai_Google_Drive__read_file_content, gsheet 1_Zz6vMPjaYXm7u53CG0cva2xEOUkzZg0tgjo7l_dDm4).

    line 45
  3. Very specific:

    Read the firm's own tracker spreadsheet (gsheet ID from the per-firm table in _shared/fd-sot-sources.md).

    line 51
  4. Very specific:

    If a tab isn't returned or a firm's tracker gsheet is unreadable, note it and proceed — this skill degrades like fd-blocker-sweep, it does …

    line 56
  5. Very specific:

    When run standalone (not via full-sweep), there may be no fresh evidence to diff against — in that case the sweep simply reports the …

    line 67
  6. Very specific:

    Who Jeff is: read _shared/fd-firms.md and paste its "Who Jeff is" block verbatim.

    line 76
  7. Distinctive:

    Never write to a Google Sheet: No tool exists to do this.

    line 142
  8. Distinctive:

    Never let Vantage write anything — Sol owns delivery, matching the established convention (Vantage/Scout return content, Sol writes).

    line 145
  9. Distinctive:

    Never silently resolve a Binder-vs-tracker disagreement.

    line 147
  10. Distinctive:

    Never treat a sparse or newly-created firm tracker (Baker Tilly, Archer Lewis) as an error.

    line 149
  11. Distinctive:

    Never batch-approve; Jeff approves the PROPOSED-CHANGES block, not a summary of it.

    line 150
  12. Distinctive:

    Never infer a row is resolved from silence in the evidence.

    line 151

≈ “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
<!-- audience: internal -->

Sweep Binder — On-Demand

Diffs "what the FD Program Binder and the firm's own tracker gsheet currently say" against "what the rest of a sweep just learned" (Jira, Slack, email, transcripts), and proposes the changes as a diff block for Jeff's approval. This skill never writes to a gsheet — no connected tool can do a cell-level Sheets write today (see the open question below) — so on approval it produces a finished, labeled markdown file for Jeff or the firm's PM to apply by hand.

Reads _shared/fd-firms.md for the current firm list and _shared/fd-sot-sources.md for every gsheet ID this skill reads — read both before running; do not hardcode IDs here.

When to run

  • Jeff says "sweep binder," "update the tracker," "reconcile [firm] tracker," or "binder sweep."
  • As part of /full-sweep (Step 5), invoked once per firm after sweep-jira and the transcript/Slack/email lanes have run, so it has fresh evidence to diff against.

Who does the synthesis — Vantage

This is closer to fd-blocker-sweep's shape than sweep-jira's: reconciling the Binder's current rows against fresh multi-source evidence (Jira, Slack, transcripts) and adjudicating disagreements (Binder wins on conflict, per the established rule) is real synthesis over a messy corpus, not a single structured query. Sol dispatches Vantage, model tier claude-opus-4-8 (high capability) — same reasoning as fd-blocker-sweep Step 3: bounded judgment over messy input, structured by a fixed adjudication rule (Binder wins), not open-ended.

Step 1 — Resolve scope

Default scope is all 5 firms (per _shared/fd-firms.md). If Jeff names one firm, scope to it. "Sweep all firms" dispatches one Vantage subagent per firm, in parallel (the fd-blocker-sweep Step 2 precedent — one unit of work per subagent).

Step 2 — Read the Binder + the firm's tracker

Per _shared/fd-sot-sources.md:

  1. Read the FD Program Binder (mcp__claude_ai_Google_Drive__read_file_content, gsheet 1_Zz6vMPjaYXm7u53CG0cva2xEOUkzZg0tgjo7l_dDm4). It flattens all tabs into one text stream — locate the Agent Tracker, Risks & Open Items, and Decisions Log tabs by their header text and extract only each tab's own rows (same technique fd-blocker-sweep and fd-onsite-availability already use for the Per-Firm Team Roster tab — stop at the next tab's header, don't let rows bleed across).
  2. Read the firm's own tracker spreadsheet (gsheet ID from the per-firm table in _shared/fd-sot-sources.md). A brand-new firm's tracker (Baker Tilly, Archer Lewis) may be sparse or empty — that is expected, not a failure; report it as "tracker not yet populated" rather than an error.

If a tab isn't returned or a firm's tracker gsheet is unreadable, note it and proceed — this skill degrades like fd-blocker-sweep, it does not refuse.

Step 3 — Accept inline evidence from other sweep lanes

This skill accepts optional inline context passed in by the calling skill (typically full-sweep): findings from the Jira sweep (sweep-jira), the Slack/Outlook recap (slack-recap), and meeting notes (process-meetings) for the same window. Paste this evidence into the Vantage spawn prompt alongside the Binder/tracker rows — Vantage's job is to diff current tracker state against this evidence, not to go re-gather it itself.

When run standalone (not via full-sweep), there may be no fresh evidence to diff against — in that case the sweep simply reports the Binder/tracker's current state for the firm with no proposed changes, and says so plainly.

Step 4 — Dispatch Vantage

Spawn Vantage per firm. The spawn prompt must be fully self-contained (no session memory). Include:

Who Jeff is: read _shared/fd-firms.md and paste its "Who Jeff is" block verbatim.

Identity and grounding discipline (paste verbatim):

You are Vantage, reconciling the FD Program Binder and a firm's tracker gsheet against fresh sweep evidence. You never fabricate. Every factual claim about current tracker state cites the Binder or tracker tab and row; every proposed change cites the evidence it rests on (which lane, which source, what date). Anything you infer is tagged [Unverified]. You end with the trace footer specified below.

Adjudication rule (paste verbatim):

When the Binder and the firm's own tracker disagree on the same fact, the Binder wins for agent status/execution — surface the disagreement as a data-quality item, do not silently pick one. When fresh sweep evidence (Jira, Slack, transcript) disagrees with what the Binder currently shows, that is exactly the case this sweep exists to catch — propose the update, cite the evidence, and let Jeff decide; do not silently update your own read of "current state."

Output template Vantage fills — a PROPOSED-CHANGES block per firm, one entry per divergence:

  • NEW — a row/fact the fresh evidence surfaces that isn't in the Binder or tracker yet (e.g. a new agent, a new blocker). Full proposed row content, with the source and date.
  • UPDATE — an existing row whose evidence changed (status moved, new owner). Show the current cell value, then the proposed value, with the source note and date.
  • RESOLVED-CANDIDATE — a Binder/tracker row that fresh evidence suggests is now closed or done. Show the evidence; Jeff confirms, Vantage never marks it resolved itself.

End with the trace footer:

---
**Sources:** [Binder tab + row, tracker tab + row, and each evidence source cited]
**Claims marked [Unverified]:** [list, or "none"]

Judgment boundaries (paste verbatim):

You must NOT: edit the Binder or any tracker gsheet (no tool exists to do this — see the open gap below); decide which of a Binder/tracker disagreement is correct without surfacing it as a data-quality item; treat a sparse or empty firm tracker as an error; infer a row is resolved because the evidence went quiet (silence is "no new signal," never resolution).

Output contract: Vantage returns the full PROPOSED-CHANGES content to Sol. Vantage does not write anywhere — Sol handles delivery (Step 5).

Step 5 — Sol review + delivery (propose-then-approve, matches Jeff's decision)

Per Jeff's explicit decision (higher-risk writes go through propose-then-approve): Sol shows the PROPOSED-CHANGES block to Jeff. On Jeff's OK, Sol — not Vantage — writes the approved diff.

Known gap: no Sheets-write tool exists. The only Google Drive MCP tools available today (search_files, read_file_content, get_file_metadata, list_recent_files, create_file, copy_file) cannot edit a cell or row in an existing spreadsheet. create_file makes a new file; it cannot patch the Binder or a tracker gsheet in place. Do not invent a fake tool call or claim a write happened if it did not.

So, on Jeff's approval, Sol's fallback is: write the approved, finalized diff to owners_inbox/tracker-updates_<YYYY-MM-DD>_<firm>.md, clearly labeled as "Approved — apply by hand to the Binder / <Firm> tracker; no automated write available." State exactly which tab and row each change applies to, so applying it is mechanical. This file is the deliverable until a Sheets-write tool is connected — at that point this skill should be revisited to write directly.

Step 6 — Trace footer

Carried through from Vantage's output (Step 4) into whatever Sol delivers — the read-only current-state summary (if no changes) or the tracker-updates_<date>_<firm>.md file (if changes were approved). Never deliver either without the footer.

What this skill must NOT do

  • Never write to a Google Sheet. No tool exists to do this; the approved-diff markdown file is the fallback, not a workaround to be silently dropped once a write path exists — that transition should be a deliberate skill update, not assumed.
  • Never let Vantage write anything — Sol owns delivery, matching the established convention (Vantage/Scout return content, Sol writes).
  • Never silently resolve a Binder-vs-tracker disagreement; always surface it as a data-quality item.
  • Never treat a sparse or newly-created firm tracker (Baker Tilly, Archer Lewis) as an error.
  • Never batch-approve; Jeff approves the PROPOSED-CHANGES block, not a summary of it.
  • Never infer a row is resolved from silence in the evidence.

Related: _shared/fd-firms.md, _shared/fd-sot-sources.md, fd-blocker-sweep/SKILL.md (the Binder-wins adjudication rule this skill reuses), sweep-jira/SKILL.md (a common evidence source), full-sweep/SKILL.md (the orchestrator that invokes this for all 5 firms).

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.