Discovery Loop
Skill2 starsChanged 4 months agoMITView on GitHub
Keep generating and testing NEW hypotheses against a live domain instead of stopping at a closed verdict - observe what is actually succeeding, log a belief-state with credence, attack every confirmation streak. Activate in any standing investigation (an edge, a root cause, a market, a growth lever, a performance hunt) when a verdict doc or kill-list exists and might be read as final; when all recent evidence came from your own priors rather than observation; or when someone else is visibly succeeding at what your verdict said was impossible. Trigger signals: "we already proved that doesn't work", a rival/competitor result contradicting your conclusion, "keep looking", resuming an investigation after a closure.
In this repository38 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- Adversarial VerifySkill
- Change ControlSkill
- Code ReconnaissanceSkill
- Correction PropagationSkill
- Debugging PlaybookSkill
- Deep DecompositionSkill
- Delegation DisciplineSkill
- Disclosure Is Not A FixSkill
- Divergent IdeationSkill
- Effort CalibrationSkill
- Empirical ValidationSkill
- Error RecoverySkill
- Extract ApproachSkill
- Failure Mode AwarenessSkill
- ForesightSkill
- Frontier Workflow ModeSkill
- Human HandoffSkill
- Intent ClaritySkill
- Leverage FirstSkill
- Live State TruthSkill
- Memory HygieneSkill
- Open MandateSkill
- Output StructuringSkill
- Plan GateSkill
- Proactive RigorSkill
- Product ThinkingSkill
- Prompt EngineeringSkill
- Publish HygieneSkill
- Research MethodologySkill
- Ruthless EditorSkill
- Scope FenceSkill
- Security Pattern ReviewSkill
- Self-Improvement LoopSkill
- Session OrientationSkill
- Structured ReasoningSkill
- Verification DisciplineSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
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 58
Kill or promote, in writing: Every hypothesis resolves to a dated verdict with receipts.
- Distinctive: line 66
Every active credence can show the evidence chain that produced its current value.
- Distinctive: line 67
No credence rose ≥3 consecutive times without a logged kill-attempt.
- Distinctive: line 68
No conclusion generalizes beyond the hypotheses actually tested.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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
Discovery Loop
Purpose
Verification kills bad ideas; nothing in a kill-oriented process creates new ones. A finished investigation with an honest verdict ("no edge found", "not reproducible", "root cause is X") quietly becomes a wall: future sessions treat the kill-list as the space. The two documented failures this skill exists for: (1) a universal negative read off a finite list — "we tested 9 strategies, all dead" became "no edge exists," until an outsider's success disproved it in one evening; (2) belief moving on vibes — credence climbing through a streak of confirmations with no attack, or a "winner" anointed from a short window that a full-history audit later showed was net-negative. The loop makes generation a standing duty and belief an explicit, attackable number.
When to use this skill
- Standing investigations that survive across sessions: market/edge hunts, recurring incidents, performance regressions, growth experiments, research programs.
- A verdict/kill-list exists and is about to be cited as the reason not to look.
- External evidence contradicts your closed conclusion (a competitor does the "impossible" thing) — that is a lead, not an annoyance.
- You notice every recent hypothesis came from your own head, none from observation.
When NOT to use
- One-shot tasks with a definite end (a bug fixed and verified; a question answered).
- Domains where more hypotheses have no value (the decision is made, the system is being decommissioned).
- As a license to re-litigate a kill WITHOUT new evidence — re-opening requires a new observation, not boredom with the verdict.
The loop (every session the investigation is touched)
- Observe before theorizing. Spend the first cycle on what the domain is DOING now — who is succeeding, what changed, what the live data shows — via the cheapest real-observation channel (public fills, logs, dashboards, competitor output). Winners' observed behavior outranks your priors as a hypothesis source.
- Generate ≥1 NEW hypothesis, dated, even half-baked. From the observation, not from re-reading old brainstorms. Log it in the hypothesis ledger with a status (OPEN).
- Keep a belief-state, not a feeling. Each hypothesis carries: credence (your probability it's real, 0–1), the evidence chain (each test, dated, with the credence delta it caused), and its cheapest next falsifying test. Update the number when evidence lands — both directions.
- Attack every streak. Track consecutive credence-raising events with no kill-attempt between them. At 3, stop confirming and run the strongest attack you can design (decay slice, capacity, full-history depth, discriminating test). A streak is a signal to attack, never to size up.
- Depth before anointing. Any "winner" found in a short observation window (a performer, a config, a tactic) gets a full-history audit before it becomes evidence — short windows manufacture winners (verified: a candidate that looked strong over ~8 samples reversed hard at depth; a "most consistent" performer was net-negative over its full history).
- Kill or promote, in writing. Every hypothesis resolves to a dated verdict with receipts; killed ideas stay on the ledger (they block re-litigation without new evidence). Scope every verdict honestly: "no edge in what we generated" — never "the space is empty."
Quality bar
- The hypothesis ledger has at least one OPEN, dated entry newer than the last verdict.
- Every active credence can show the evidence chain that produced its current value.
- No credence rose ≥3 consecutive times without a logged kill-attempt.
- No conclusion generalizes beyond the hypotheses actually tested.
- New-hypothesis provenance is tracked: some fraction must come from observation, not introspection — if the operator supplies every breakthrough lead, the loop is failing.
Common failure modes
- Closure creep: the verdict doc becomes a reason to stop observing the domain.
- Introspection-only generation: brainstorming from priors while the live domain (and its winners) goes unwatched.
- Streak riding: each favorable test raises confidence and none tries to kill — the exact path to over-calling (one candidate: four confirmations ridden up, then the two hardest designed attacks broke it).
- Short-window anointing: promoting a winner/config/tactic observed briefly; survivorship at small n reads as skill.
- Credence theater: numbers logged but never moved by evidence, or moved without a recorded test.
Works with sibling skills
divergent-ideationgenerates options at a decision point; this skill makes generation recurring against a live domain with a belief ledger.empirical-validationruns each hypothesis's cheapest falsifying test (step 6's engine);adversarial-verifysupplies step 4's attacks.verification-disciplineowns the honest scoping of verdicts (step 6).extract-approachpersists resolved lessons;session-orientationcarries the ledger's location across sessions.
Provenance and maintenance
Added 2026-07-11 from a standing investigation: a closed "nothing here" verdict was scope-corrected after observation-first scanning found what the kill-list never covered; the belief-state file (credence + evidence chain + streak tripwire) then caught two would-be over-calls in its first day (an assumption retracted by measurement; an apparent top performer that dissolved under a full-history audit at scale). Re-verify by auditing standing investigations: if a session cited a verdict as a reason not to observe, or a credence rose on a streak unattacked, the loop failed — name the step.
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.

