agentleFS
BI-BlitzerBI-Blitzer/tandem-tokens/.claude/skills/harvest/SKILL.md

harvest

Skill1 starsChanged 4 months agoMITView on GitHub

Collect finished ollamaq jobs — read their compact digests, then run the Opus/Sonnet reduce/clean pass over only the residue the local model got wrong, and write the final artifact. Use after /offload jobs reach the "done" state.

17 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use handoff.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • MicrosoftSkill

    handoff

    Compact the current conversation into a handoff document for another agent to pick up.

    Also in 4 projects that use this

The file

We split this skill into the rules it gives an agent

3 rules. 0 are common, 0 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 14 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

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

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

    Clean pass (all good): if the digest shows high success and the samples look right, the local output is the deliverable.

    line 26
  2. Very specific:

    Fix the residue: if the digest shows a systematic local error (e.g.

    line 29
  3. Very specific:

    Per-row review (last resort): only read all of merged.jsonl into context when the task genuinely needs Claude to inspect every row.

    line 33

“Common” is measured: kept by at least 6 in 10 of the 14 edited versions of this file we could read. “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

/harvest — reduce and finalize offloaded work

The map phase already ran locally for $0. Harvest is where Claude tokens get spent — so spend them only on the residue, guided by each digest. See [CLAUDE.md] for the token-economy contract.

1. Find finished jobs

ollamaq status — shows counts per state and the ids in each. Note the ids in done (and failed). For one job: ollamaq status <job_id>.

2. Read the digest FIRST (cheap)

ollamaq digest <job_id> — prints results/<job_id>/digest.md (≤ 8 KB): success rate, sample outputs, the full list of failures, and the estimated cloud cost avoided. This is usually all you need to decide what to do next. Do not read merged.jsonl yet.

3. Decide the reduce strategy

  • Clean pass (all good): if the digest shows high success and the samples look right, the local output is the deliverable. Point the user at results/<job_id>/merged.jsonl (and items/). No token-heavy pass needed.
  • Fix the residue: if the digest shows a systematic local error (e.g. a format the small model didn't follow) or lists failures, prefer a mechanical fix — write a small script/transform that corrects that specific error across merged.jsonl, and run it. This keeps the reduce at O(digest), not O(n).
  • Per-row review (last resort): only read all of merged.jsonl into context when the task genuinely needs Claude to inspect every row. This is the expensive path — say so, and confirm it's warranted.
  • Escalate failures: re-run just the failed_items.jsonl items as a new small job on a stronger local model, or handle them inline.

4. Write the final artifact

Produce the cleaned/aggregated output the user asked for (a file under their project, not under results/). State plainly what you did: how many items came from the local run untouched, how many you corrected and how, and how many failed. Include the digest's "cost avoided" figure so the savings are visible.

5. gen_test / pipeline jobs

  • gen_test: the digest is a scored leaderboard — the deliverable is the top candidate(s). Report the best and its score; no reduce pass usually needed.
  • pipeline: the digest shows a stage funnel; merged.jsonl holds each item at its final stage. Verify the funnel matches intent (right counts continued vs finalized) before finalizing.

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.