proposal-forge
Skill1 starsChanged 11 days agoApache-2.0View on GitHub
Forge a research proposal from an existing matrix and gap notes, then stress-test it through attack and review rounds. Use when the task is to write or harden a research design — assemble evidence per theme, define reporting protocols and audit designs, run an adversarial review that lists fatal/major/minor objections with an explicit survival verdict, and record repair items in execution order. Input: matrices, gap notes, and the library's contradiction register. Output: a proposal document with protocol, audit table and risk/novelty-check lists, plus attack and review layers that never overwrite the original verdicts. Claim adjudication is delegated to evidence-audit, the single allowed cross-skill reference. It does not write paper prose and does not build cards or matrices.
In this repository7 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- evidence-auditSkill
- literature-radarSkill
- paper-cardSkill
- research-opsSkill
- synthesis-matrixSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 64
Do not interleave grades: fixing a P-3 while a P-1 is open produces a document that looks repaired and is not.
≈ “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
proposal-forge
A proposal is a claim about what should be done next, defended by evidence about what is already known. This skill builds that claim, then tries to destroy it before a reviewer does.
Two phases, and they must not be run in one pass: forge writes the design, attack tries to break it. A design written while being defended acquires the shape of its defence.
Phase 1 — forge
| Step | Output | Refuses to proceed when |
| 1 — assemble evidence | one block per theme: the cards, the values, the register entries | a claim has no card behind it |
| 2 — state the gap | the gap this proposal closes, quoting its gap note | the gap cannot name a decision it would change |
| 3 — commit to a reporting protocol | R1–R4: per-experiment table, aggregation and weights, negative control arm, inference statement | any of the four is left to "we will see" |
| 4 — design the audit | one row per claim: what would falsify it, what data exists, what it would cost to close | a row has no falsifier |
R1–R4 are commitments made before the experiment, not reporting conventions chosen afterwards. A proposal that cannot state its negative control arm has not decided what counts as success.
→ references/protocol.md
→ references/audit-design.md
Phase 2 — attack
Three identities, in this order, because each one requires the previous to have failed:
| Identity | Asks | If it succeeds |
| fatal | does the project, in its current form, still hold? | stop — repair or abandon, no point listing minor issues |
| major | does the central claim survive its best counter-argument? | the claim is rewritten, not patched |
| minor | what would a careful reader still object to? | collect, do not act yet |
Then the competitor route: if I were building the competing paper, what two steps would end this one? Write both steps out. If either is available, the project has a structural problem, not a presentation problem.
Survival verdict is mandatory. For every claim: survives / survives narrowed / falls. A list of objections with no verdict is a reading note, not an attack.
→ references/attack.md
Repair, in order
Repairs are graded P-1 to P-4 and executed in that order:
| Grade | Meaning | Blocks |
| P-1 | the claim cannot stand as written | everything downstream |
| P-2 | a definition or scope is ambiguous enough to change the verdict | the claim's wording |
| P-3 | a number needs a targeted re-read | that number's use |
| P-4 | structural — the design's shape is wrong | nothing until P-1..P-3 are done |
Do not interleave grades: fixing a P-3 while a P-1 is open produces a document that looks repaired and is not.
Naming note: R1–R4 (reporting commitments) and P-1–P-4 (repair order) are different systems with different scopes. They are never converted into each other, and neither is a blocker level.
Layers never overwrite
An attack pass adds a section. A review pass adds another. The earlier verdicts stay visible, and each later pass names what it overturns and why.
A document whose three review rounds look unanimous is usually a document where each round edited the last.
Challenge mode
For an adversarial argument-construction mode with round thresholds instead of a single attack pass, see references/challenge-mode.md. It is a mode of this skill, not a separate pipeline — and it is deliberately gated, because running it on a project with fewer than three supporting papers produces confident prose about an empty foundation.
Routing
| Task | Go to |
| write the reporting commitments | references/protocol.md |
| design the audit table | references/audit-design.md |
| attack a draft, grade the repairs | references/attack.md |
| run multi-round adversarial construction | references/challenge-mode.md |
| see worked cases of each failure | references/case-studies.md |
| rule on whether a claim holds | evidence-audit (the only cross-skill reference) |
| a matrix or gap note this depends on | synthesis-matrix |
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.