heavy
Skill4 starsChanged 5 months agoMITView on GitHub
Launch a pre-engineered "heavy prompt" from your local library in the current interactive session — big self-contained research/build/optimize jobs that write line-addressable structured output. /heavy lists the library; /heavy NN launches or resumes run NN.
In this repository18 other files it gives its agents.
- biz-validateSkill
- braindumpSkill
- catch-upSkill
- clone-websiteSkill
- double-checkSkill
- explain-for-grandmaSkill
- heavy-buildSkill
- heavy-reviewSkill
- make-promptSkill
- monitor-claudeSkill
- orientSkill
- overnightSkill
- reelSkill
- sanity-checkSkill
- save-sessionSkill
- shipSkill
- what-breaksSkill
- yt-tranSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 3 are distinctive, and 12 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
12
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 26
No argument: read heavy/library/INDEX.md, print the catalog (ID, Title, Goal, Cost), and ask the user which to run (this is the ONLY place …
- Distinctive: line 35
If none matches, list the catalog and stop.
- Very specific: line 36
Read its frontmatter: id, output_dir, phases, subagents, websearch_max, est_cost.
- Very specific: line 37
If subagents: false (the default), you MUST do all work yourself in this session — do not spawn the Agent tool or any sub-workflow.
- Very specific: line 42
Run folder is the prompt's output_dir (e.g.
- Very specific: line 44
Read it, note completed_steps and current_step, append a resume event to heavy/state/runs.ndjson, and continue execution skipping completed …
- Very specific: line 47
Else (new run): create output_dir, write the initial STATUS.json:
- Very specific: line 61
Create 00-summary.md with the opening `
yamlblock (status: in_progress), append a launch event to heavy/state/runs.ndjson, add an … - Distinctive: line 72
Follow the prompt body's PHASES/STEPS in order.
- Very specific: line 79
Commit after: add the step id to completed_steps, set last_commit, then git add heavy/ && git commit -m "heavy(NN): <step id>" (scoped — …
- Very specific: line 85
Never: use AskUserQuestion. When a decision is unclear, take the documented default and record it under <!-- SEC:assumptions --> in …
- Very specific: line 89
Run /compact at ~70% context.
- Distinctive: line 94
When the prompt's DONE criteria are met:
- Very specific: line 97
Update the run's row in heavy/output/INDEX.md to done with a one-line summary + folder link.
- Very specific: line 107
All run state lives under the run folder (STATUS.json) and heavy/state/ — there is no other state store to keep in sync.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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.
Files
Heavy-prompt launcher
Runs one big self-contained prompt from your heavy/library/ in this interactive session (so it
uses your interactive subscription, not metered/headless API billing). A heavy prompt is meant to be
launched and left — so once a prompt is running, do not stop to ask the user anything.
The library, outputs, and conventions live in your project repo under heavy/ (you can also keep it
at ~/.claude/heavy/). Read heavy/README.md once for the full model. Paths below are relative to
your repo root (the current working directory):
heavy/library/— the heavy prompts (NN-slug.md) +INDEX.md+_TEMPLATE.mdheavy/output/— one run folder per launch, plusoutput/INDEX.mdheavy/state/runs.ndjson— append-only event log of launches/resumes/completions
The argument $ARGUMENTS is the heavy-prompt number NN (may be empty).
Step 1 — Route on the argument
- No argument: read
heavy/library/INDEX.md, print the catalog (ID, Title, Goal, Cost), and ask the user which to run (this is the ONLY place a question is allowed — the human is present at launch). Then stop and wait. Do not auto-pick. NNgiven: proceed unattended through the steps below and never useAskUserQuestionagain for the rest of the run.
Step 2 — Load the prompt
- Resolve the library file:
heavy/library/NN-*.md(the file whose stem starts withNN-). If none matches, list the catalog and stop. - Read its frontmatter:
id,output_dir,phases,subagents,websearch_max,est_cost. - If
subagents: false(the default), you MUST do all work yourself in this session — do not spawn the Agent tool or any sub-workflow.
Step 3 — Resolve the run folder (resume or init)
- Run folder is the prompt's
output_dir(e.g.heavy/output/01-some-study/). - If
output_dir/STATUS.jsonexists andstatus == "in_progress": this is a RESUME. Read it, notecompleted_stepsandcurrent_step, append aresumeevent toheavy/state/runs.ndjson, and continue execution skipping completed steps. Reuse this same folder — never create a new dated one. Else (new run): create
output_dir, write the initialSTATUS.json:{ "prompt_id": "NN-slug", "started": "<ISO8601>", "phases": [...from frontmatter...], "current_phase_index": 0, "current_step": null, "completed_steps": [], "last_commit": null, "lock": {"started": "<ISO8601>"}, "status": "in_progress" }Create
00-summary.mdwith the openingyamlblock (status: in_progress), append alaunchevent toheavy/state/runs.ndjson, add anin_progressrow toheavy/output/INDEX.md, thengit add heavy/ && git commit(the commit is the first checkpoint). (Get the timestamp fromdate -u +%Y-%m-%dT%H:%M:%SZ— do not invent times.)Never
git add -A— other concurrent work may have unrelated uncommitted changes, and-Awould sweep them into the heavy commit. Everything a run writes lives underheavy/, sogit add heavy/is the correct, complete, scoped stage.
Step 4 — Execute the phases
Follow the prompt body's PHASES/STEPS in order. For each step:
- Advance before work: set
current_stepand (if entering a new phase) bumpcurrent_phase_indexinSTATUS.json. - Do the step, writing output as anchored, append-only sections
(
<!-- SEC:id -->…<!-- /SEC:id -->) into the prompt's specified files. Never rewrite a completed earlier section; only append new ones (re-running a partial step may rewrite that step's own anchored block). - Commit after: add the step id to
completed_steps, setlast_commit, thengit add heavy/ && git commit -m "heavy(NN): <step id>"(scoped — never-A). This is the durable checkpoint. - At the end of each phase, regenerate
MAP.md(table:| id | title | file | lines | ~tokens | one-liner |) and append aphase_doneevent toheavy/state/runs.ndjson.
Discipline while running:
- Never use
AskUserQuestion. When a decision is unclear, take the documented default and record it under<!-- SEC:assumptions -->in00-summary.md. - WebSearch ≤
websearch_maxfor the whole run. - Respect the prompt's GUARDRAILS (security, low-risk-edits-only, verify external facts).
- Run
/compactat ~70% context. If the session is cut off, the next/heavy NNresumes fromSTATUS.json.
Step 5 — Finish cleanly
When the prompt's DONE criteria are met:
- Finalize
00-summary.md(set yamlstatus: done), regenerateMAP.md. - Set
STATUS.jsonstatus: "done", clearlock. - Update the run's row in
heavy/output/INDEX.mdtodonewith a one-line summary + folder link. - Append a
doneevent toheavy/state/runs.ndjson. - Final
git add heavy/ && git commit -m "heavy(NN): done"(scoped — never-A). - Print a short banner: prompt id, status, output folder, top findings, how many commits.
- Stop. A heavy run is one bounded job — do not loop or relaunch.
Notes
- This is launched in an interactive session on purpose (your interactive subscription, not metered API billing). It is human-launched every time — do not schedule it to fake a human.
- All run state lives under the run folder (
STATUS.json) andheavy/state/— there is no other state store to keep in sync.
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
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.

