Anti-patterns

• Find the symptom you see, then follow the link to the fix
• Principles list theirs under “Signal of Violation”; patterns and stack pages under “Anti-patterns”
• Each linked section has the full list; this index keeps the most common

Correction Loops

Symptom	Anti-pattern	Fix
Same correction, again and again	Steering by chat correction	Correction Diagnosis
Twenty turns of “no, not like that”	Prescribing before stating constraints	Problem Before Prescription
Solved the wrong problem	No problem statement or constraints; verifying first, framing never	Problem Before Prescription, Correction Diagnosis, Cause Chain
Wandered mid-task	No plan review or milestone	Spec, Then Build, Cause Chain
No check can say the work is done	“Make it work” as the only criterion	Spec, Then Build
Prompt grows while misses recur	Tuning the prompt over an inconsistent codebase; adding context to fix a context problem	Consistency as Leverage, Correction Diagnosis, Memory & Context
The same miss on the second attempt	Re-rolling a systematic miss	Fail Fast, Recover Smart
Refinement never ends	No stopping criteria	Iterative Refinement
Each pass reverses the last one	Passes that undo each other’s work	Iterative Refinement
Wrong format or style	No exemplar to mirror	Few-Shot Examples, Structured Output, Cause Chain
Shallow answers on hard steps	Tier too small for the step	Chain of Thought, Step-Level Routing, Cause Chain
Model over-fits to specifics	Examples that are too similar	Few-Shot Examples
Diminishing returns; context waste	Too many examples	Few-Shot Examples

Context and Memory

Symptom	Anti-pattern	Fix
Context fills; signal gets buried	Dumping whole files or logs “just in case”	Memory & Context
Ignored a repo convention	Convention not written down, or inconsistent	Consistency as Leverage, Cause Chain
Output got worse after material was added	Irrelevant or conflicting context	Memory & Context, Cause Chain
Long run drifts or loses the thread	Automatic summary as the only memory	Context Handoff, Progress Breadcrumbs
Confident answers about your own data that are wrong	Relying on what the model memorized	RAG
A constraint is lost across compaction or handoff, and the next model or session breaks it	Treating compaction as free; a handoff summary that drops a constraint	Memory & Context, Context Handoff, Step-Level Routing
Output quality falls as the session ages	Never resetting a drifting session	Correction Diagnosis
Same fix made in two sessions	Correction never captured	Discovery Propagation, Cause Chain
Model uses content that does not apply	Retrieving too much or irrelevant content	RAG
Agent cites code that has moved	Stale code index	Repository Context
Wrong tool picked; context fills	Too many tools loaded at once	Tool Integration

Review and Verification

Symptom	Anti-pattern	Fix
Review finds nothing the author missed	Same agent and context generates and reviews	Adversarial Review
Critique praises or lists generic tweaks	“Is this good?” as the critique prompt	Self-Critique
A “tests pass” claim that cannot be checked, or that hides a failing exit code	“Tests pass” with no command or output; trusting subagent summaries over artifacts	Reviewable Output, Subagent Fanout, Agent Architecture
Diffs too large to review	Refactor mixed with behavior change; task too large for one brief	Reviewable Output, Cause Chain
Green checks, broken behavior	The loop edits the test until it passes	Verification Loops
Plausible code that does not run	No runnable check	Verification Loops, Cause Chain
Valid JSON with wrong values	Treating schema-valid as correct	Structured Output
The scaffold keeps emitting the old shape	Scaffold drifts from the target schema	Mechanical Scaffolding
The two copies disagree over time	Validation duplicated in prompt and code	Mechanical Scaffolding
A check nobody has seen fail	Gates that always pass	Pipeline Orchestration
Agreement accepted without reasons	Using consensus to avoid thinking	Multi-Model Consensus
Conclusions that do not follow ship	Not reading the reasoning	Chain of Thought
Bad data flows downstream	Trusting tool output without validation	Tool Integration

Delegation and Orchestration

Symptom	Anti-pattern	Fix
Architecture corrected in review	Delegating the architecture decision	Delegation Fit
Fixing output takes longer than writing it	Wrong work delegated	Delegation Fit, Cause Chain
The diff grows past the task	A brief with no non-goals	Delegation Fit
Edits collide	Parallel agents sharing a branch	Subagent Fanout, Agent Architecture
Coordination context runs out	Orchestrator doing the work itself	Agent Architecture
The prompt grows with every new task	One mega-agent instead of decomposition	Agent Architecture
A late failure restarts the whole pipeline	No resumability	Pipeline Orchestration
Agents spawned for work one thread finishes sooner	Fanout by habit	Subagent Fanout

Long Runs

Symptom	Anti-pattern	Fix
The run never knows it is done	No stop condition	Unattended Runs
One request burns tokens and returns nothing	Uncapped retries	Unattended Runs, Fail Fast, Recover Smart
The human babysits an unattended run	Checking in every ten minutes	Unattended Runs
Nobody can tell where the run stands	Status only in chat scrollback	Progress Breadcrumbs
The next session redoes or breaks it	Half-implemented feature with no note	Context Handoff
Progress file says done; behavior is not	Agent rewrites acceptance tests	Context Handoff
The first outage is when a fallback gets written	Failure handling added after launch	Fail Fast, Recover Smart

Human Checkpoints

Symptom	Anti-pattern	Fix
Sign-off depends on who ran the agent	Undefined human boundary	Human in the Loop
Near-100% approval in seconds	Rubber-stamping	Human in the Loop, Checkpoint Gates
Agent took an irreversible action unasked	Approval requested by prompt text only	Checkpoint Gates
Reviewer rebuilds context to decide	A card with no recommendation	Checkpoint Gates
Work idles in the approval queue	Blocking the whole run on one decision	Checkpoint Gates, Human in the Loop
External effects with no recorded approval	Outward-facing actions with no approval step	Tool Integration

Models and Routing

Symptom	Anti-pattern	Fix
Cost explodes on simple tasks	Biggest model for everything	Task Routing, Model Selection
Quality collapses on hard tasks	Cheapest model for everything	Task Routing
Short, hard steps get a weak model	Routing on input length	Task Routing, Step-Level Routing
Cold caches on every call	Switching models every step	Step-Level Routing
An outage stops the work	No fallback model	Task Routing, Model Selection
A regression nobody logged	Floating model alias	Model Selection, Reproducibility

Testing and Evaluation

Symptom	Anti-pattern	Fix
A prompt change broke something that worked	No baseline	Prompt Regression Testing, Evaluation & Benchmarking
Tests fail on rephrasing alone	Exact string matching on free prose	Prompt Regression Testing, Reproducibility
Flaky tests get retried or deleted	Single runs on nondeterministic output	Prompt Regression Testing, Reproducibility
A bug “fixed” because a rerun passed	Debugging by regenerating	Reproducibility
The metric rises; the work does not improve	Trusting a score without reading transcripts	Evaluation & Benchmarking
Real use breaks what tests passed	Testing on synthetic examples only	Dogfooding
Generation fails silently	Not testing retrieval	RAG

Governance and Currency

Symptom	Anti-pattern	Fix
The same convention argued each review	Settled questions reopened	Authority Cascade
Work stalls when they rotate off	Conventions held in one person’s head	Authority Cascade
Skills tuned for a retired model still load	Drift	Deliberate Currency
Tooling reopened on every release	Churn	Deliberate Currency
Nobody knows why a clause is there	Prompt archaeology	Skills & Prompts, Memory & Context
No sweep has deleted a skill	A skill library that only grows	Skills & Prompts
Shared instructions change unnoticed	Silent mutations without review	Discovery Propagation
Skills and patterns lost in the move	Switching harnesses without migrating skills	Harness Selection

Cost and Attention

Symptom	Anti-pattern	Fix
Bill falls; rework and steering rise	Optimizing raw token spend	TTV, Cost Management
Total cost per outcome rises	Optimizing per-call price	Step-Level Routing
Nobody can state the cost of a win	Tokens logged with no outcome ID	Observability & Logging, TTV
Spend surprises with no breakdown	Ignoring cost until the bill arrives	Cost Management
Limits stop work that mattered	Hard limits on critical paths	Cost Management
Failures visible in logs, unnoticed	Traces nobody reads	Observability & Logging