Skip to content

method-map.md

Latest commit

 

History

History
28 lines (23 loc) · 2.27 KB

method-map.md

File metadata and controls

28 lines (23 loc) · 2.27 KB
Raw

Method map: failure → fix → artifact

Don't fix agent reliability by dumping more text into one instruction file. When you see a specific failure, add the smallest artifact that addresses it. This table maps the common failure modes to their fix and the file in this suite that implements it.

Failure mode What it looks like Primary fix Artifact
Cold-start confusion A new session rediscovers setup/status from scratch Make the repo the system of record progress.md (+ harness-init)
Scope sprawl Starts many features, finishes none; touches unrelated files Restrict active scope to one feature feature_list.json (WIP=1)
Premature completion Claims "done" with no runnable proof Bind completion to recorded evidence clean-state-checklist.md, the verify path
Fragile startup Every session re-learns how to boot the project Standardize setup + verification init.ps1 / init.sh
Weak handoff Next session can't tell what's verified, broken, or next End with an explicit handoff session-handoff.md (via session-wrap)
Lost rationale Agent undoes a deliberate choice after a reset Persist the "why" outside chat decisions.md
Context rot Quality drops as the window fills with stale text/noise Route + progressively disclose; reset over compaction AGENTS.md as a map, docs/, context-management.md
Subjective review "Looks good to me" review that depends on taste/memory Score against fixed categories /review-gate prompt-command
No clean exit Repo decays; next session spends 30 min on rescue Make clean state part of "done" clean-state-checklist.md (via session-wrap)

How to use this

  1. Notice the actual symptom (not a vague "the agent is bad").
  2. Find the matching row.
  3. Add or strengthen that one artifact — most are already stamped by harness-init; the fix is usually using it consistently via the session-start / session-wrap rituals.
  4. Re-check after a couple of sessions. Treat the lowest-performing subsystem as the bottleneck, but confirm with real session outcomes before declaring victory.

See harness-model.md for the subsystem each artifact belongs to.