agentleFS
frenzymathfrenzymath/Danus/.agents/skills/elaboration/SKILL.md

elaboration

Skill391 starsChanged 3 months agoApache-2.0View on GitHub

Write a high-signal mathematical synthesis from global memory and the fact graph for the Codex main agent's own strategy and worker dispatch.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 3 are distinctive, and 2 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.

  • 2

    Very specific

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

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

    Read only the shared stores — never a worker's private local memory (a layer boundary, and the reason this is cleaner than a log-scraping …

    line 46
  2. Distinctive:

    Do not skip an interface even if its row is trivial — flag trivial matches so a cold reader knows they were considered.

    line 157
  3. Distinctive:

    No numerical distance estimates: Never "≈ 8–12 facts", "2–4 focused runs", "80% done".

    line 200
  4. Distinctive:

    No process telemetry: No worker counts, service state, mtimes, run scheduling — this is purely mathematical.

    line 202
  5. Very specific:

    Literature-aware: Before presenting a route as novel or committing heavily to it, use search_arxiv_theorems broadly with varied …

    line 213

≈ “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.

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

Elaboration

You are the main agent. At each strategic cycle, including the global review on each roughly 30-minute heartbeat, you distill the project's current state into one elaboration when the synthesis materially changes: a readable, deeply analytical synthesis for your own reasoning and worker dispatch. The heartbeat itself always requires a fresh global appraisal, but it need not create a duplicate elaboration when nothing material changed. You may ask speculative Codex subagents to explore individual gaps or undertake sustained technical reasoning, but their reports remain unverified and must be labelled as hypotheses. You author the next master_guidance yourself.

The elaboration is also what you draw on to keep the operator informed.

Template invariants (validate every elaboration against these)

A well-formed elaboration satisfies all of the following — a worker or linter can check them mechanically:

  • Five sections, in order, no dropped heading: §0 Mathematical verdict · §1 Closed components and obsolete routes · §2 Interface contract table · §3 Dangerous heuristic lines and strategies not to pursue · §4 Missing bridge lemmas. An empty section is written as its honest empty-state line, never omitted.
  • The two fixed empty-state lines (§1) are used verbatim when a subsection is empty:
    • _(no signed-closed components yet)_
    • _(no failed or obsolete routes recorded; absence here does not mean the strategy is unique)_
  • Exactly seven status labels, UPPERCASE, drawn only from: CLOSED · SUBSTANTIAL · PARTIAL · DANGEROUS · FALSE AS STATED · OBSOLETE · UNKNOWN.
  • §0 opens with exactly one bolded verdict line and contains the status dashboard table + the sub-task status summary table + the approach portfolio + the Current best proof skeleton + the Central missing lemma.
  • Every fact_id cited exists in the fact graph; no invented ids, no paraphrase substituted for a verified statement.
  • Published via gm_add(kind="elaboration", …) with verifiable left at its default (false).

Input Contract

Read only the shared stores — never a worker's private local memory (a layer boundary, and the reason this is cleaner than a log-scraping summary agent). All reads are project-scoped for the main agent (project=<p>):

  • global memory — findings, dead ends, recent verification traces, the current master_guidance. Read via gm_search, or as a fallback by reading the raw runtime/projects/<p>/global_memory/<kind>.jsonl files.
  • fact graph — the verified facts, their statements, and the DAG. Read via fact_search, or as a fallback by reading the raw runtime/projects/<p>/fact_graph/facts/*.md files: what is established vs. still open, and how facts compose.
  • the project's problem statement — the fixed goal and, if present, its enumerated sub-tasks / intended proof architecture.

The fixed goal is sacred

Quote the goal and do not change or weaken it — do not redefine, simplify, restrict to a special case, or substitute an easier proxy. If the evidence suggests the goal may be false or unreachable by the current strategy, say so plainly while keeping the goal fixed.

Template — five sections

Produce one markdown document with these sections, in order. Omit a section's body only by writing the honest empty-state line, never by dropping the heading.

0. Mathematical verdict

Open with one of these, in bold on its own line:

Not solved. … | Counterexample found. … | Verified complete proof. … | Solved. …

Then:

  • Closed components — what is signed-closed today (1–2 sentences; cite fact_ids).
  • Viable proof architecture — one sentence naming the current best route.
  • Main blocker — what concretely blocks right now (1–2 sentences; cite fact_ids).
  • Highest-priority unresolved bridge — the most leveraged missing lemma / integration package.
  • Method failure vs. proposition failure — state explicitly whether the evidence indicates a method has failed (the conjecture may still hold) or the proposition itself may be false. Use the phrase "method failure" or "proposition failure" verbatim.
  • Calibration caveat — one line warning the reader against over-reading status labels (e.g. "Do not read SUBSTANTIAL/CONDITIONAL as 'almost solved' — every such row has an unmatched hypothesis on the actual model.").

Then a status dashboard (one table) with at least these rows: Fixed goal (UNCHANGED, with goal text); Verified complete proof (YES/NO); Verified counterexample (YES/NO); Signed-closed sub-tasks (count + names); Main blocker (a specific lemma, not vague); Routes marked false/obsolete (YES/NO + which); Highest-priority unresolved task (P0/P1/P2 with the exact mathematical task).

Then a sub-task status summary (one table: Sub-task | Status | Closed facts | Conditional facts | Main missing interface), one row per sub-task the problem enumerates. Use only these UPPERCASE labels:

  • CLOSED — verified on the actual construction, no remaining hypothesis-matching. A theorem import or conditional package being available is not CLOSED — that is SUBSTANTIAL. CLOSED is rare; default away from it.
  • SUBSTANTIAL — a conditional package exists, but ≥1 input/output hypothesis is unmatched on the actual construction. The default for a sub-task with load-bearing tools not yet applied to the actual model.
  • PARTIAL — isolated ingredients only; no coherent conditional package yet.
  • DANGEROUS — a plausible shortcut that is false / insufficient / hypothesis-sensitive.
  • FALSE AS STATED — a once-plausible formulation now refuted; do not pursue as stated.
  • OBSOLETE — superseded by a better route; do not pursue.
  • UNKNOWN — insufficient verified information.

Strict CLOSED test. For each sub-task you are tempted to mark CLOSED, ask: "Is there a verified fact that handles this on the actual construction, with zero remaining hypothesis to match?" If you cannot answer YES with a specific fact_id and zero remaining work, mark SUBSTANTIAL. Over-marking CLOSED is the single most damaging error here — it reads as "no further work needed."

Then an approach portfolio (one table: Approach | Mechanism | Mathematical frontier | Decisive obstacle | Evidence for/against | Active/parked | Revisit condition). Include every credible route still worth remembering, not only the currently dominant route. Preserve parked routes and their return conditions so that recent work cannot silently erase a serious alternative. If a major route choice has changed, state the alternatives considered and the mathematical reason for the change; this decision must also be preserved in the subsequent master_guidance.

End §0 with Current best proof skeleton (6–12 short numbered lines: the smallest structure that closes the goal if the central missing lemma were known, with fact_ids where facts apply) and Central missing lemma (the single most precise unresolved statement, at full precision — all quantifiers, definitions inlined for self-containment, and one short "why this is non-trivial" paragraph if warranted).

1. Closed components and obsolete routes

  • Signed-closed components — a bullet list ("<math content> — fact_ids …") or the line _(no signed-closed components yet)_.
  • Failed or obsolete routes — a table (Route | FALSE AS STATED / OBSOLETE | one-line reason citing a fact_id / concrete obstruction), or _(no failed or obsolete routes recorded; absence here does not mean the strategy is unique)_. FALSE AS STATED = a plausible reduction now refuted; OBSOLETE = superseded by a simpler live route.

2. Interface contract table

The single most important diagnostic — a human reader uses it to find exactly which input/output hypothesis is unmatched on the actual model. For each interface in the proof architecture (use the exact sub-task names the problem enumerates: per-stage A/B/C…, each transition B→C, C→D…, and the meta-reduction to the original statement):

<Interface name> — <one-line role in the proof>

Input required. <precise mathematical conditions step i+1 demands of step i's output — normality, Q-factoriality, R-Cartierness, dimension, …; not just "compatibility"> Output claimed. <what the existing conditional package guarantees, conditional on its own hypotheses> Available facts. fact_id — one-line statement; … Missing verification on the actual model. <numbered: the specific hypothesis-matches not yet carried out on the actual construction> Failure mode if ignored. <one or two sentences: what concretely breaks downstream — e.g. "If K_W+B_W is not Q-Cartier, 'by negativity lemma' is vacuous and the crepancy conclusion is unjustified."> Status. <one of the seven labels>

Do not skip an interface even if its row is trivial — flag trivial matches so a cold reader knows they were considered. Inline the definitions of load-bearing terms so a cold reader need not consult the problem statement. If the problem is built around a single central lemma rather than a pipeline, produce one interface row for the central reduction in the same format. Strict CLOSED rule applies per row: if "Missing verification" is non-empty, the status is SUBSTANTIAL or weaker — never CLOSED.

3. Dangerous heuristic lines and strategies not to pursue

  • Dangerous heuristic lines — 3–8 specific shortcut statements found verbatim or in close paraphrase in the problem's strategy outline or prior notes (preserve the wording so the reader can locate them). For each: a Status (one of: "Not automatic" / "Mathematically incorrect as stated" / "Conditional only" / "Method shortcut, not a proof step" / "Conflated with a stronger claim") and one 2–4 sentence paragraph on the precise reason it is unjustified, citing fact_ids; give the correct rephrasing where one is needed.
  • Strategies not to pursue — 4–8 one-line anti-routes, each concrete: "Do not <specific action>: <one-line reason>." Vague entries ("avoid weakness") are forbidden.

4. Missing bridge lemmas

Identify 2–4 bridge lemmas — concrete intermediate facts that, if proven, would either prove the central missing lemma or unblock a specific interface row. For each:

Bridge k — <short name>

Target statement. <full, precise mathematical statement> Prerequisites. <conditions the inputs must satisfy for the statement to type-check> Existing facts to use. fact_id — one-line claim; … (only ids that actually exist) Missing checks. <numbered: the specific hypothesis-matches / sub-proofs still needed> Closure criterion. <one paragraph naming the exact proof obligation that closes this bridge>

Order by leverage: Bridge 1 unblocks the most downstream / has the highest payoff per unit effort. State for each whether it is independent (parallelizable) or dependent — this is what lets you put different workers on different bridges.

Discipline (hard rules)

  • Do not treat conjectural material as verified — a claim without a fact_id is awareness, not truth.
  • No numerical distance estimates. Never "≈ 8–12 facts", "2–4 focused runs", "80% done". Distance is qualitative. Estimating a numeric distance is forbidden.
  • No process telemetry. No worker counts, service state, mtimes, run scheduling — this is purely mathematical. Do not discuss how the elaboration was produced, scheduled, or delivered.
  • No agent-facing directives. The elaboration is a synthesis, not a dispatch order. Use it afterward to author master_guidance and worker assignments.
  • Honest, not reassuring. Surface hidden assumptions, possibly-false statements, and places where a status label may mislead. Do not round SUBSTANTIAL up to "almost done".
  • Global, not locally captive. Judge the whole portfolio and the route to the fixed goal. Fact volume, proof length, and activity inside the primary route do not by themselves show macro-level progress.
  • Literature-aware. Before presenting a route as novel or committing heavily to it, use search_arxiv_theorems broadly with varied formulations and technique names. Record a concise technique map in global memory: mechanisms, exact assumptions, limitations, relevant arXiv identifiers/results, and possible interfaces with this problem. Understand and adapt established strategies before inventing new machinery; literature notes are not facts.
  • Four-hour macro audit. At least once every four hours of active work, explicitly reassess and record the full approach portfolio, mathematical frontier and obstacle of each route, evidence for/against it, worker allocation, and whether to continue, complement, park, or resume each route.

Output Contract

Publish the elaboration to global memory with gm_add:

  • kind: elaboration
  • claim: the §0 verdict line (the bolded opener + the one-line main blocker)
  • evidence: the full five-section markdown body
  • links: {"fact_ids": ["…", "…"]} — the facts you cited (only ids that exist in the fact graph)
  • (verifiable defaults to false for this kind — it is a synthesis/judgment, not an objectively checkable claim; leave it unset.)

Then reason over the elaboration yourself. Optionally give precise pieces to exploratory subagents, label their reports unverified, synthesize the result into master_guidance, and dispatch Danus workers afterward.

Tools

Reference the role=main MCP tools by name (never internal engine paths):

  • gm_search / read runtime/projects/<p>/global_memory/<kind>.jsonl — gather findings, dead ends, recent verifications, current master_guidance.
  • fact_search / read runtime/projects/<p>/fact_graph/facts/*.md — the verified facts and the DAG (fact_search to pull the facts bearing on a sub-task; read the files for the full statements/proofs and predecessor structure).
  • gm_add (kind elaboration) — publish the synthesis.
  • search_arxiv_theorems — use repeatedly with varied formulations and technique names to map the relevant literature, understand established mechanisms and hypotheses, and check whether missing bridges or nearby results already exist.

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.