re-feedback

Experience Feedback

Complete Workflow

1. Collection (three sources unified):
   • Session retrospective: pitfalls/new methods from this analysis session (already ready when called in step 4 of [[re-analyze]])
   • Article scan:node bin/wxsource.mjs Fetch Kanxue/WeChat articles (node bin/wxsource.mjs kanxue thread <帖子ID> --md / node bin/wxsource.mjs wechat <文章URL> --md), distill each article
   • Manual input: a user's one-liner/note
2. Distillation and desensitization: per references/intake.md —— pitfall format **标题**：现象——…；原因——…；对策——…, desensitize at the same time
3. Domain classification:references/intake.md decision table → grep skill library as fallback → if still uncertain, ask the user (do not guess)
4. Three-tier handling (neutral presentation, no recommendation labels): First ask the user for the tier: 'How should this experience be handled? 1. Publish issue 2. Local ingestion 3. No ingestion', then after selection execute the corresponding tier; do not presuppose a default tier
   • ① Publish issue:which gh + gh auth status probe; if available then gh issue create --repo dslsdzc/rev-skills --label 经验, title 'Experience: <skill>: <one-liner>', body per references/issue-template.md; if unavailable → output a markdown draft for manual submission; label reports nonexistent → remove --label and retry or prompt the user to create it first; repository unreachable → output draft
   • ② Local ingestion: grep for deduplication → append to that skill's references/experience.md(create if absent, entries include source + date) → SKILL.md 'Common Pitfalls and Traps' keeps selections and appends an experience.md link (create the file first, then add the link, to avoid broken links) → npm test validate → commit per the step 5 agreement (or ask the user first)
   • ③ No ingestion: confirm and end
5. (article scan path optional) Commit commit: 'Experience: round N scan ingested X articles——<article list>', message lists each skill's newly added entries

When to take which path (decision tree)

• Input = summary of this analysis session → session retrospective source; goal = accumulate this real-world session
• Input = article URL / post ID / article text → article scan source; distill each article, skip tool/promotional posts per intake.md standards
• Input = user's one-liner/note → manual source; distill directly
• Three-tier selection (neutral presentation): 1. Publish issue (publicly shareable, reusable by others) 2. Local ingestion (accumulated in this repository, published with the repository) 3. No ingestion (retained only within the session)

Cross-domain joint

• [[re-analyze]] step 4 (optional) calls this gateway for session retrospective
• Domain classification targets all skills ([[re-binary-core]] [[re-malware]] [[re-firmware]] [[re-protocol]] [[re-mobile]] [[re-anti-analysis]] [[re-cracking]] [[re-vuln]] [[re-ctf]] [[re-managed]] [[re-forensics]] and their atomic skills)
• Article fetching uses repository bin/wxsource.mjs; after ingestion use npm test(validate.mjs) to validate structure

Common Pitfalls and Traps

• Distilling without desensitizing and directly ingesting/publishing issue → leaks project identity —— red line 2; before ingestion list proper nouns item by item for self-check
• Presentation with 'most recommended' etc. strongly pushing → hijacks the user's judgment —— red line 1; only list options neutrally
• Guessing when domain classification is uncertain → experience is written under the wrong skill and cannot be retrieved next time —— ask the user
• Not running npm test after committing → broken links/format errors break CI — after every commit you must npm test
• Duplicate entries committed repeatedly → repository bloat — grep keywords to deduplicate before committing
• Article scanning does not skip tool/promotional posts → experience devalued — filter according to the intake.md skip criteria
• Assert preconditions before trusting tool output: Phenomenon — the tool appears to run normally but produces empty/loose/no-op output (missing official runtime returns empty output, hooks get zero hits when the architecture does not match, repackaged with a packer shows a black screen on a real device, screenshot is blank when the screen is not woken); Cause — tools all have hard preconditions (official runtime/native architecture/not packed/environment ready), and when these are violated they do not report an error but fail silently; Countermeasure — before accepting any output, explicitly assert that its preconditions hold on that target, and always treat silence and empty results as suspicious first; if three structurally distinct direct paths all fail, it is blocked by a runtime invariant, so stop retrying and switch environments/abstraction layers/targets (Source: reverse-skills (inliver233), MIT)
• Write multi-agent collaboration state to a handoff document: Phenomenon — parallel multi-agent analysis cannot reconcile, dead ends are revisited, hypotheses are repeatedly verified; Cause — state exists only in chat history, and agents have no common baseline for what is "confirmed/excluded"; Countermeasure — a single controller is the sole verification authority (agent hypotheses must be reproduced on a device by the controller to count as confirmed), each agent has one goal per round, and the operation site/confirmed facts (with verification method)/required excluded list/suggested first step are written to the handoff document; reconciliation is sorted by evidence category (runtime > static > hypothesis) (Source: reverse-skills (inliver233), MIT)
• Distillation wording triggers collaboration model refusal: Phenomenon — distillation artifacts containing words such as "crack/bypass" are refused in the collaboration process; Cause — trigger words are intercepted by content policy, unrelated to content quality; Countermeasure — during the distillation stage, systematically rewrite into engineering language (code recovery/compatibility adaptation/integrity verification adaptation/runtime signature adaptation), and append an authorization premise statement (own code recovery/interoperability/security audit/CTF/education, not authorizing action against protected software), presented as neutral and from the same source as red line 1, and self-check accordingly before committing (Source: reverse-skills (inliver233), MIT)
• For more experience, see this skill's [[experience]] (complete experience collection)