show-your-work
Skill1 starsChanged 3 months agoMITView on GitHub
Use when the user asks to ship the Show Your Work weekly podcast — explains frontier-lab AI alignment research (Anthropic, OpenAI incl. its misalignment reports, Google DeepMind) to newcomers in a two-voice Explainer/Skeptic episode, published to its own public RSS feed via syw_gather / syw_script_plan / syw_write and render.py. Skips the standard production interview because defaults are pre-set.
6 other projects use a version of this.
Adopted by
Most recently taken up by
- youcisla/Agent-Foundrytheir own version, first seen Nov 16, 2026
- dheerajjha/show-your-worktheir own version, first seen Oct 11, 2026
- quatico-solutions/agent-skillstheir own version, first seen Jul 2, 2026
- ahirmaan/show-your-worktheir own version, first seen Jun 25, 2026
- muyuhill/100books-to-skillstheir own version, first seen Jun 11, 2026
- ChewbaccaRoars/pathfinder-skillstheir own version, first seen Mar 22, 2026
Other versions6 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 6
- muyuhill/100books-to-skills★ 14same name
- dheerajjha/show-your-work★ 4same name
- ahirmaan/show-your-work★ 2same name
- ChewbaccaRoars/pathfinder-skills★ 2same name
- quatico-solutions/agent-skills★ 2same name
- youcisla/Agent-Foundry★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- daily-podcastSkill
- frontier-commitsSkill
- surface-tensionSkill
- tts-evalSkill
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 2 are distinctive, and 4 are very specific to this project.
0
Common
Most of the 5 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 14
Every cold open discloses it.
- Very specific: line 16
Every pushback line carries a basis that must be the post's own stated limitations or a matched independent check.
- Very specific: line 80
Every pushback traces to the post's stated limitations or a matched independent source (enforced: basis).
- Distinctive: line 81
Never call a lab "better" or "worse" at safety.
- Distinctive: line 87
Never point the listener at a link or the show notes.
- Very specific: line 123
Follow each block, save the result as {"ok": true, "lines": [{"speaker": ..., "text": ...}]} and run python3 S/syw_write.py accept …
“Common” is measured: kept by at least 6 in 10 of the 5 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.
Files
Show Your Work
A weekly podcast that explains frontier-lab alignment research to people who don't read papers. Each episode is one feature (one post, five scenes), up to four short briefs, and a sign-off. Two voices: the explainer states each claim the way the post states it, always attributed; the skeptic asks the newcomer's question, then pushes back.
The show is written and voiced by Claude, and Anthropic is one of the labs it covers. Three mechanisms answer that conflict, and none is optional:
- Every cold open discloses it.
syw_write.validate_framerefuses a cold open without "Claude" and "Anthropic". - An Anthropic feature gets a fixed reminder line. The assembler prepends
ANTHROPIC_REMINDER; no writer writes it. - Every pushback line carries a
basisthat must be the post's own stated limitations or a matched independent check.validate_featurerefuses anything else.
RSS-first. The show ships through render.py's web-only mode on its own feed. The R2 publish is the ship. save-to-spotify is never invoked.
Design: docs/superpowers/specs/2026-09-26-show-your-work-design.md.
Layout
./syw_gather.py: sources → items;gather/seed/commit. Metadata only. It drops the OpenAI alignment index'sopenai.comcross-posts, because the writers' WebFetch gets 403 there, and logs each one todropped.jsonl(#245)../syw_script_plan.py: the deterministic plan: feature pick, briefs, casebook, checks, rotations../syw_write.py: prompt filling (fill), output validation (accept), manifest + beats assembly (assemble)../prompts/digest.md: a one-body fact sheet for a check or a casebook incident../prompts/write_feature.md: the five-scene feature. One body plus check digests../prompts/write_brief.md: one single-post brief. One body../prompts/write_casebook.md: the casebook brief. No body, digests only../prompts/weekly.md: a stub pointing here.
The renderer is the sibling skills/daily-podcast/render.py, used unchanged.
Episode shape
Cold open (with disclosure) → feature (5 chapters) → 0–4 briefs (research briefs, then at most one casebook) → sign-off (with the assigned button). About 13–18 minutes, 7–11 chapters.
The feature arc
| # | Slot | Job |
| 1 | hook | The assigned opening move; why a newcomer should care. |
| 2 | method | What the researchers did, with every term of art defined on first use. |
| 3 | finding | What they found, attributed to them. |
| 4 | pushback | The skeptic's objections, each traced to a basis. |
| 5 | stakes | Why it matters, and what is still open. |
Assigned variety
Seeded by week_index(date). The lengths are pairwise coprime except fourth_wall × first_speaker, which never share a scene.
| Rotation | Length | Values |
intro_mode | 3 | question, moment, ledger |
fourth_wall | 4 | maker, conflict, method, standard |
opening_move | 5 | excerpt, number, scenario, question, claim |
first_speaker | 2 | explainer, skeptic |
button | 7 | grader, reward, transcript, spec, eval, sandbox, homework |
The daily show's example lines (FALLBACK_BUTTONS, FALLBACK_FOURTH_WALLS) are burned here: a frame containing one is refused.
Visual beats
Beats are written now and drawn by the (separate) video stage. They live in <workdir>/beats.json, never in the manifest. A beat that fails its guard is dropped and logged; its line still ships. If accept's own fetch of the post fails, it is logged (stage: fetch) and every number, chart, quote and transcript beat drops as post text unavailable; the scene still validates on its structure.
| Type | Guard |
number | the value appears in the post text |
chart | every y value appears in the post text, never read off a figure |
quote | ≤25 words, verbatim (typography-blind) |
transcript | verbatim, ≤600 chars, role cot / tool_call / tool_result / model / user |
diagram | ≤6 nodes; edges join known ids |
term | defined once per episode |
Editorial rules
- The explainer attributes every claim; a finding is never stated as settled fact.
- Every pushback traces to the post's stated limitations or a matched independent source (enforced:
basis). - Never call a lab "better" or "worse" at safety.
- Define every term of art on first use.
- A misalignment report is told as what the lab says happened. Add no drama.
- Quotes and transcript excerpts are verbatim (enforced).
- Charts use only numbers stated in the post text (enforced).
- Disclose in every cold open; a fixed reminder opens any Anthropic feature (enforced).
- Never point the listener at a link or the show notes.
Manifest
syw_write.assemble_manifest writes these keys; the collision test in tests/test_syw_write.py keeps them clear of every other feed.
{
"ship_mode": "web",
"show_name": "Show Your Work",
"r2_manifest_name": "manifest-show-your-work.json",
"r2_key_prefix": "show-your-work/",
"slug_prefix": "syw-week-of"
}The cast is recorded clips, each cloned on the base model as {ref_audio, ref_text} (the clip and its .txt transcript) — never presets, which the production Qwen3 checkpoint does not have (#243):
explainer:skills/daily-podcast/refs/house_voice.wavskeptic:skills/show-your-work/refs/skeptic.wav(Surface Tension's Ethan clip pitched down 1.5 semitones; seesyw_write.CAST_CLIPS)
Unattended weekly run
This section is the canonical procedure for shipping an episode with no human in the loop. It is the single home: a scheduler invokes this skill and follows this section, never carries its own copy (prompts/weekly.md is a stub pointing here, and a drift test keeps it one). Be decisive, don't ask clarifying questions, and if you cannot proceed, print one FAILED <reason> line and stop.
Let D = today (YYYY-MM-DD), S = this skill's directory (${CLAUDE_PLUGIN_ROOT}/skills/show-your-work/ when set; otherwise the path this SKILL.md was loaded from), <root> = the plugin root, the parent of skills/ (two levels above S), and W = $TMPDIR/daily-podcast-show-your-work-<D> (explicit, so render.py never auto-deletes it; the daily-podcast- prefix lets --prune-workdirs sweep it later).
- Gather.
python3 S/syw_gather.py gather --date D --out W/candidates.json. OnGATHER FAILED, printFAILED <that reason>and stop. - Plan.
python3 S/syw_script_plan.py plan --date D --candidates W/candidates.json --out W/plan.json. OnPLAN skip <reason>, printSKIPPED <reason>and stop — no filler episodes (PLAN skip already shippedmeans an earlier run already committed this week's plan). OnPLAN FAILED <reason>, printFAILED <reason>and stop. - Digest the checks and the casebook. For each URL in
plan.checksand each item in acasebookbrief:python3 S/syw_write.py fill digest --workdir W --url <url>→ run that prompt in its own subagent context (one body per context, nothing else in it) → save the subagent's final message toW/out/digest-<n>.txt→python3 S/syw_write.py accept digest --workdir W --output W/out/digest-<n>.txt --url <url>. A refused digest is dropped: a missing check narrows the skeptic's basis topost-limitations; a missing incident leaves the casebook. - Write the feature.
python3 S/syw_write.py fill feature --workdir W→ one subagent context → save →python3 S/syw_write.py accept feature --workdir W --output <file>. OnACCEPT refused, retry once with a fresh subagent. If it is refused again, re-plan without it:python3 S/syw_script_plan.py plan --date D --candidates W/candidates.json --out W/plan.json --exclude <feature url>, then redo steps 3–4 on the new plan (a new plan deletesW/writes/, since every accepted write belongs to the plan it was written for; digests are kept). A second failed feature isFAILED feature refused twice. A run where every writer came backAUTHis a credential failure, not a thin week — report it as such. - Write the briefs. For each index
iofplan.briefs:python3 S/syw_write.py fill brief --workdir W --index i→ one subagent context → save →python3 S/syw_write.py accept brief --workdir W --output <file> --index i. A refused brief is dropped; the episode still ships. A casebook whosefillis refused (FILL refused casebook has no digests— every incident digest in step 3 was refused) is dropped: skip its subagent. Write the frames in the main context, after step 5 (the instructions list only the briefs that will air):
- the cold open:
python3 S/syw_write.py fill cold_open --workdir Wprints its instructions — the assignedintro_modeandfourth_wallangle TEXTS, the episode's titles, the 400–800 band, the disclosure rule (ONE dry sentence: written and voiced by Claude, a model made by Anthropic, one of the labs covered) and the burned lines; - the sign-off:
python3 S/syw_write.py fill sign_off --workdir Wprints its instructions — the assignedbuttonTEXT (a thanks and ONE dry joke, worded fresh), the 350–500 band and the burned lines.
Follow each block, save the result as
{"ok": true, "lines": [{"speaker": ..., "text": ...}]}and runpython3 S/syw_write.py accept cold_open|sign_off --workdir W --output <file>. Rewrite until accepted.- the cold open:
- Assemble.
python3 S/syw_write.py assemble --workdir W --summary "<one sentence on this week's feature>"→W/manifest.json,W/beats.jsonandW/aired.json(the feature URL plus each brief item that went into the manifest — a refused brief, or a casebook incident whose digest was refused, is not in it). It uses a brief's write only if it was accepted for this plan's brief at that index, and refuses (ASSEMBLE FAILED) a feature or frame written for another plan. OnASSEMBLE FAILED already published …, an earlier run published this episode fromWand died before step 9: skip step 8 (never re-render — the slug's R2 objects are immutable-cached), run step 9 against the existingW/render.log, and report as step 10 does, takingbriefsfromW/manifest.json(its segments minus 7). On any otherASSEMBLE FAILED <reason>, printFAILED <reason>and stop. - Render in the background.
python3 <root>/skills/daily-podcast/render.py --manifest W/manifest.json --workdir W > W/render.log 2>&1withrun_in_background, and monitorW/render.log— the 10-minute foreground Bash cap kills a long render. Never pass--dry-run(this is a real episode) and never pass--skip-preflight. - Commit, only after
render.pyhas exited (the background task finished, not just a quiet log): the log's last JSON object is the resultcommitparses, and a render still running has not printed it yet.python3 S/syw_gather.py commit --plan W/plan.json --render-output W/render.log. It readsW/aired.json(beside the plan) and marks exactly those stories covered — never the whole plan, so a dropped brief and every leftover return to the pool — and only if the render's final JSON saysweb-readyandpublished. If that final JSON is notweb-ready+published, the episode did not ship: reportFAILED <reason>. If it is, the episode is live even whencommitwas refused or failed (e.g. it ran before the render exited): re-runcommit— re-running is the recovery; never re-render or re-publish. - Report once.
SHIPPED <mp3_url> feature=<url> briefs=<n> gather_errors=<n> r2=ok—mp3_urlfrom the renderer's final JSON,feature=plan.feature.url,briefs= the brief chapters that aired (the ASSEMBLE line'ssegmentsminus 7),gather_errors= the GATHER line'serrors(sources that failed this week);SKIPPED <reason>; orFAILED <reason>.
Setup
mkdir -p ~/.config/show-your-work && echo '{}' > ~/.config/show-your-work/config.json— every key has a default (syw_gather.DEFAULT_CONFIG); override only what you need.python3 skills/show-your-work/syw_gather.py seed --date <today>— marks everything currently published as seen so episode one covers only what arrives next.gatherrefuses until this has run (GATHER FAILED no seen.json — run seed first (SKILL.md Setup)). Seed refuses (SEED FAILED <adapter errors>, nothing written toseen.json) if any lead source failed; re-run it until it printsSEED ok.- R2 credentials as for every show (
render.py --selftest).
State (~/.config/show-your-work/)
| File | Written by | Contract |
config.json | human | {} is valid; unknown keys refuse |
seen.json | seed, commit | URL → {date, role}; only after a verified ship (or seed), and only what aired.json lists; corrupt refuses |
observed.json | gather, seed | URL → first date parsed; month-precision items age from it |
features.jsonl | commit | one row per shipped feature; drives the lab and kind penalties |
dropped.jsonl | gather, accept | failed adapters, refused writes and digests, failed validator fetches and dropped beats; observability only |
render.py also writes the shared ~/.config/daily-podcast/covered.json for every chapter's source_url, so the daily show will skip a lab post this show already covered. Accepted; this show reads only seen.json.
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.

