Failure Mode Awareness
Skill2 starsChanged 4 months agoMITView on GitHub
Identify what could go wrong before it goes wrong. Activate when designing anything new, before committing to a plan or recommendation, before shipping changes with real blast radius, and when evaluating ideas, strategies, or proposals. Trigger signals: "will this work?", "review this plan", "any risks?", designing an integration or workflow, writing a strategy or proposal, or a plan that contains zero mention of what could fail. Complements adversarial-verify, which attacks finished work; this skill runs earlier, on designs and plans.
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
- Discovery LoopSkill
- Divergent IdeationSkill
- Effort CalibrationSkill
- Empirical ValidationSkill
- Error RecoverySkill
- Extract ApproachSkill
- 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
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
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 15
Before recommending a course of action the user will invest in.
- Distinctive: line 16
When reviewing someone else's plan, proposal, or idea.
- Distinctive: line 17
Before shipping anything with real blast radius: data migrations, external sends, public artifacts.
- Distinctive: line 55
No definition of done, so it's never done.
- Distinctive: line 68
Every High-impact mode has a mitigation, a watch signal, or an explicit acceptance.
≈ “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
Failure Mode Awareness
Purpose
Cheap foresight beats expensive hindsight. Most failures aren't exotic — they're standard, catalogued, and visible in advance to anyone who asks "how does this break?" before asking "how does this work?" This skill runs that question early, when changing course costs a sentence instead of a rewrite.
When to use this skill
- At plan time (
plan-gaterisk section) and design time — before the approach hardens. - Before recommending a course of action the user will invest in.
- When reviewing someone else's plan, proposal, or idea.
- Before shipping anything with real blast radius: data migrations, external sends, public artifacts.
When NOT to use this skill
- On finished work — that's
adversarial-verify's job (attack the artifact, not the idea). - On low-stakes, reversible tasks. Enumerating failure modes for a typo fix is noise.
- As a reason not to act: the output is mitigations and watch-items, not paralysis.
Operating procedure
Step 1 — Run the premortem. "It's three months later and this failed. What's the most boring, most likely reason?" Write down the top 2–3. The boring qualifier matters: real failures are usually mundane (nobody used it, the data was dirty, the dependency changed), not dramatic.
Step 2 — Sweep the relevant catalog (pick the sections that apply):
Software failures
- Edge inputs: empty, null, zero, negative, huge, duplicate, malformed, unicode.
- State: concurrent access, partial failure mid-operation, retries causing double-execution, stale caches.
- Integration: the other system is down, slow, rate-limiting, or returns a shape you didn't expect.
- Environment: works locally, differs in prod (versions, permissions, paths, env vars, clock, locale).
- Scale and time: fine at 10 rows, dead at 10 million; fine today, broken at month-end/DST/leap year.
Reasoning failures
- Anchoring on the first hypothesis; confirmation bias in which evidence gets collected.
- Survivorship bias (studying only what succeeded); base-rate neglect ("this time is different").
- Extrapolating from an unrepresentative sample; correlation read as cause.
Communication failures
- The lede is buried; the reader acts on the first paragraph and misses the correction in the fourth.
- Wrong audience register: jargon for executives, hand-waving for engineers.
- Ambiguity someone will resolve the wrong way ("next Friday", "the new version", unowned action items).
Business/strategy failures
- Demand assumed, not validated; distribution unplanned ("build it and they will come").
- Unit economics that don't survive scale; single points of dependency (one customer, one channel, one platform's policy).
- Competitor/platform response ignored; timeline assumes nothing slips.
Workflow failures
- Handoffs where context evaporates; work that silently drifts from the request (see
scope-fence). - No definition of done, so it's never done; blocked items with no owner, waiting forever.
- Verification skipped under deadline pressure — precisely when errors are most likely.
Step 3 — Rate and respond. For each identified mode: likelihood (L/M/H) × impact (L/M/H). Then choose:
- Mitigate now (high L or H impact, cheap to address) — change the design, add the check.
- Watch — name the early-warning signal that says it's happening.
- Accept — explicitly, with a note. Accepted risks are fine; unexamined ones are not.
Step 4 — Fold results into the plan. Failure modes that survive go into plan-gate's Risks section or the deliverable's caveats — not into a separate document nobody reads.
Quality bar
- The top failure mode identified is plausible and specific to this task, not generic ("something might break").
- Every High-impact mode has a mitigation, a watch signal, or an explicit acceptance.
- The whole pass fits the stakes: three lines for a Medium task, a structured table only when Critical.
Common failure modes (of this skill itself)
- Catastrophizing: listing twenty theoretical risks, drowning the two real ones. Rank ruthlessly; report the top few.
- Generic risk lists: "scope creep, technical debt, timeline risk" pasted onto any project. If the risk list would fit a different project unchanged, it's decoration.
- Premortem as ritual: writing risks, then designing exactly as if they didn't exist. Step 4 is the point.
- Only-software blindness: checking edge cases in code while the actual failure is that the output will be misread (communication) or unneeded (strategy).
Example
Plan: "Nightly script exports user data to the analytics vendor."
Premortem top hits: (1) Script silently stops running and nobody notices for weeks — boring, likely. Mitigate: success heartbeat + alert on absence. (2) Schema change upstream breaks the export mid-quarter — watch: row-count drop alert. (3) PII lands in the export that shouldn't — high impact: mitigate now with an explicit allowlist of fields rather than select *. Accepted: vendor outage delays data by a day (low impact, retry handles it).
Works with sibling skills
- Feeds the Risks section of
plan-gate; runs during design, before execution. adversarial-verifyis the post-hoc twin: this skill predicts failure in plans, that one hunts it in finished work.proactive-rigorsurfaces the missing information class of risk; this skill covers the known-mechanism class.debugging-playbookuses the software catalog in reverse — as a source of hypotheses for existing failures.
Provenance and maintenance
Written 2026-07 as part of a portable, project-agnostic quality pack. Catalog entries are standard, well-documented failure classes, not project claims. Re-verify by post-incident review: when something fails that this skill should have caught, add the mode to the catalog section it belongs to. Prune any entry that has never once fired for your work.
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.

