forge-blog
forge-blog is a set of blog skills for an agency team. A keyword goes in and a writer's brief comes out; a draft goes in and one checklist comes out. Each skill says plainly what it checked and what it could not, and none of them invents experience the author does not have.
Last updated
How to read this
Every skill goes through the same ten steps. Every skill starts folded shut; click its name to open it. A finished step is folded shut; open it to see its history. A skipped step counts as done only if it was completed later. Each mark has a text label, so colour is never the only signal.
- Done
- Running
- Failing
- Skipped or out of order
- Not started
- Started
- Progress
- Failed
- Fixed
- Done
- Skipped
Skills
- blog-brief In progress
- blog-check In progress
- blog-draft In progress
- blog-refresh In progress
- blog-plan In progress
- blog-repurpose In progress
33 of 60 steps done · 55%
blog-brief
In progress, 4 of 10 steps Takes a keyword and writes a brief: who ranks, what no ranking page says yet, and what to cover. It stops if it cannot find something new to add.
Totals appear when the skill ships
4 of 10 steps done · 40%
-
1 Research (Scout) Done
-
Scout started a deep dive into existing blog plugins, to borrow good ideas and credit them.
-
Scout's deep dive is done: 8 public projects reviewed, 25 ideas ready to copy with credit.
-
Scout finished the research without calling in any helper. Alex re-checked the licence of the four main sources.
-
Skill 1 research started. Scout is on it.
-
-
2 Plan (Rook) Done
-
Rook finished round 1 of the rework spec: an argument step, a request for the author's own proof, article length set from competitors, and a 12-point usefulness rubric.
-
Rook finished round 2 of the rework spec: 7 playbooks by article type, a context template with a worked example, and a rubric per type.
-
Rook started the spec for the rebuilt blog-brief (version 2).
-
Rook wrote a follow-up spec: outline skeletons per article type, source rules, and slots for first-hand experience.
-
Rook's plan finished.
-
Rook started the skill plan.
-
-
3 Merge and check (Alex) Done
-
Alex reviewed the rework spec. It leans too much on scripts, and has no context template or per-type playbooks yet. Sent back for round 2.
-
Alex reviewed round 2 and it is ready to build. No leak from the test material found.
-
Alex checked the plan and fixed one hand-off: blog-draft does not exist until skill 3.
-
-
4 Board yes on plan Done
-
Board approved the plan. Build could start.
-
-
5 Build (Kit), Windows and Linux tests Running
-
The blog-brief rework is merged on the main build branch. 1,038 tests pass. A cross-platform review comes next.
-
Kit fixed the fault that dropped trend figures in the blog-brief rework.
-
Kit finished the blog-brief rework on its separate branch. 801 tests pass. 29 tests from the first version were removed, each with a stated reason. Kit now fixes a fault that dropped trend figures, then builds the rework for blog-draft and blog-check.
-
The first run of the blog-brief rework was cut off by too much command output. The skill, seven playbooks, scripts and tests were written, but 32 tests fail and 10 error. No trial run yet.
-
A second run started to fix the tests and commit. A trial run on a real topic follows.
-
Kit started building the blog-brief rework on a separate branch.
-
Windows and Linux CI green. 507 tests pass.
-
Board decided Windows is the main platform and Windows CI must pass before anything merges.
-
Two blockers fixed. 293 tests pass.
-
Rule added: an angle anyone could copy does not count as new information. 233 tests pass.
-
Added a log of which commands really ran, so a report the model wrote by hand cannot pass as a script result. 195 tests pass.
-
Kit fixed the 9 failures. Alex re-ran: 144 of 144 tests pass.
-
First build run was cut off by too much command output. 135 of 144 tests passed; 9 failing.
-
Kit started the build.
-
-
6 Review (Sable) Running
-
Kit fixed all six blockers and two smaller points. The test suite passes with 1,067 tests. Sable is re-checking.
-
Sable's second check left one blocker, a status note in one file. It is fixed and the tests still pass.
-
Sable reviewed the reworked blog-brief and said fix first. Six blockers: an old readme, a write boundary, a figure check when search data is missing, the date of the age check, front matter on merged outputs, and a stray folder in the tree. The tests still pass. Kit is fixing them now.
-
Last two blockers fixed. Sable's review passed for the final test runs.
-
Sable review of the extra brief sections: fix first (a summary figure matched by substring, unsourced claims, a fake volume stamp).
-
Sable re-checked the pilot fixes: 4 of 4 confirmed.
-
Sable got round the anti-fake check. Kit hardening it.
-
Sable round 4: a check report existed although the script never ran, so the skill had written it by hand. Kit adds an anti-fake check.
-
Sable said ready for the Board. Branch pushed for the Board to read.
-
Sable confirmed the 7 fixes and found one more must-fix in the link checker.
-
Kit fixed all 7. Alex re-ran: 158 of 158 tests pass.
-
Sable found 3 must-fix problems (the link checker could be redirected, unstamped facts, no per-field sourcing) and 4 smaller ones.
-
Sable's review started.
-
-
7 Quality test, skill alone Running
-
The judge was still failing the reference brief on formatting rules alone, so those rules were loosened. A second calibration run followed.
-
With the formatting rules gone, the reference brief fails on only 3 real content flaws: two instructions that contradict each other, a figure that does not match the data, and placeholder text left in. The judge is now calibrated. The bar for the rework is the reference brief with those 3 flaws fixed.
-
The new judge for blog-brief says the strong reference brief beats our first version in all 4 comparisons. The old judge had said we won 3 of 4. The first calibration check failed because the reference brief had been shortened. The full text is now used.
-
A trial brief from the rework came close to the strong reference brief. Alex's judge passed every key point except one: trend figures were dropped because a safety check was too strict. That is a fault in the plumbing and is being fixed. Vera is writing judge prompts for all six skills and will check them against the reference brief and the first version.
-
Direction change for blog-brief. Scripts now handle plumbing: fetching, counting, formatting and hard safety checks. The model does the thinking. The rebuild is judged on how useful the brief is against a strong reference, not on pass counts. We call it the blog-brief rework.
-
The first real-world brief was not good enough when held against a strong reference brief. The earlier head-to-head test judged the wrong thing. blog-brief will be rebuilt on the reference structure, keeping our own checks. The real draft is on hold.
-
Diagnosis of the last run: most main-suite judge failures were faults in the test rubrics, now fixed. One test case was wrong. Three cases failed because the test setup did not allow web search. Two faults in the skill were fixed and tested. One real habit remains: outlines assumed events the author had not described. Kit is fixing that.
-
Kit fixed the outline habit. Outlines now use only experience the author has given. 734 tests pass.
-
New run of all 17 cases started, with web search allowed in the three cases that need it.
-
Regression run on the updated test cases: 4 of 17 passed. The main-suite judges fail as they did in earlier rounds, and the skill stops without a brief in three cases. The new experience-question case also fails. Vera is diagnosing; no more paid runs until there is a clear fix.
-
A regression run was voided by a fault in the test setup: the sandbox never started. All 17 cases are being re-run.
-
Kit fixed the three faults from the first real-world brief. The skill now asks the author for first-hand experience. Output goes to a place the sandbox allows. The checker accepts a valid volume stamp. 727 tests pass and the attack checks still hold.
-
Test cases updated to cover the new experience question, with one new case that tests it directly. A paid re-run of the affected cases has started.
-
First real-world brief written for a live topic. It showed three faults: the skill never asked the author for first-hand experience, one step wrote output where the sandbox refused it, and the checker rejected a valid volume stamp. Kit will fix them.
-
Round 10 (03:40): 1 of 12 cases passed. The new section judges failed in the same way every time, so this was not noise. Causes found: some skill wording overstated claims, and some test rubrics were wrong. Both fixed. Paid re-run on 4 Nov.
-
Test redesign: the two whole-brief judges become five small judges, one per section. Size cap removed; a repetition check replaces it.
-
Round 9: 2 of 12 cases passed. Score up, but the honesty and fit-to-send judges still fail most cases. Rook is finding the root cause before more fixes.
-
Kit got the judge's reasons for one case and fixed the causes. Round 9 is running.
-
Round 8: 2 of 12 cases passed. Honesty labels still fail in 9 cases and fit-to-send in 7. Briefs are 11,000 to 16,000 characters, too long. Kit fixing.
-
Round 7: 1 of 12 cases passed. Same pattern as before.
-
Final runs across three models: the rule checks pass widely, but the quality judges fail almost every case. Not shippable on that evidence.
-
Pilot 3: every case wrote a brief and ran its checks. The skill beat no skill in 8 of 8 comparisons. 3 of 9 cases passed the fit-to-send judge; most other fails were over-literal checks or faults in the test case.
-
Pilot 1: 1 of 11 cases passed. Some briefs were never written and the check step never ran. A sandbox fault made script steps invalid. Alex reported it late.
-
First quality pilot started: 11 cases, each run with and without the skill.
-
-
8 Quality test, all skills together Running
Needs skill 3 (blog-draft) so the chain brief, draft, check can run end to end.
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
Checklist due on the morning of 3 Nov.
-
Our side of the build is finished. Work is paused until the Mac run results come back.
-
A new pack for the Mac run is ready. It adds a guard that stops a new post clashing with one already on the site. The test suite passes with 1,080 tests, and the pack has been validated.
-
The pack for the Mac run is ready and has been sent to the Board: the plugin, a sample site and a checklist.
-
Board: one full run of all six skills on his Mac, once all six are ready (00:16).
-
-
10 Board final yes, then main Skipped or out of order, not completed
Not yet proven: the quality test is still failing.
-
blog-brief v0.1.0 published and merged to main.
-
Board said ship, so blog-brief went to main before its quality test passed. Out of order.
-
blog-check
In progress, 6 of 10 steps Takes a draft and returns one checklist: fix before sending, worth fixing, polish. It checks banned words, secrets, links, claims and voice, and never edits the draft.
Totals appear when the skill ships
6 of 10 steps done · 60%
-
1 Research (Scout) Skipped or out of order, completed
-
The comparison with industry tools was done after the build, not before. Out of order, but complete. It added 5 gaps to the plan.
-
-
2 Plan (Rook) Done
-
Rook finished the rework spec for blog-check. It gives a verdict of ship, fix, block or undecided instead of a score.
-
Rook's plan for blog-check saved.
-
Skill 2 started after the Board said to move on.
-
-
3 Merge and check (Alex) Done
-
Alex accepted Rook's choices: no partial drafts and no 0-100 total score.
-
Alex folded the 5 industry gaps into the plan.
-
-
4 Board yes on plan Skipped or out of order, not completed
-
No separate yes on this plan. The Board said to move on to skill 2.
-
-
5 Build (Kit), Windows and Linux tests Done
-
Kit built the blog-check rework. 841 tests pass across the build. One hygiene check still fails.
-
On main. Windows and Linux CI green.
-
Windows CI green on the fix branch.
-
Windows CI: 2 failures left, both file path separators.
-
Five industry gaps added. 402 tests pass.
-
Built: 394 tests pass, 10 eval cases written.
-
-
6 Review (Sable) Done
-
After the owner's first real use, Kit rebuilt the checks. Sable reviewed the changes 10 times and found real gaps each time; all are fixed. 822 tests pass.
-
All gaps from the review are fixed in the code. Every known attack is now blocked. Sable's re-review of this skill is next.
-
Review found gaps: an incomplete review could still read as clear. Fixing.
-
Independent review by Sable started (this step was skipped earlier).
-
Sable's review was skipped before the paid test run. Still to do.
-
-
7 Quality test, skill alone Done
-
Live test on a made-up post: the first check found all 4 planted faults with no false alarms. After the owner's reply, the re-check said Ready to publish with no new points.
-
Second planted-fault trial on blog-check: it caught 6 of the 7 faults that earlier runs missed. Verdict: fix.
-
Planted-fault trial on a faulty draft: scripts alone caught 7 of 18 faults. After a fix to how author facts are traced, it caught 12 of 18 faults and 8 of 9 blockers, with no false alarms. One veto case was missed.
-
Round 6: 11 of 11 pass, score 1.00, 0.55 better than no skill.
-
Round 5: still 7 of 11. All rule checks pass; 4 cases fail only the judge on whether the checklist is useful.
-
Round 4: 7 of 11 cases passed, 0.52 better than no skill. Fails: a false fix (case 1), a keep-list phrase flagged (case 6), and a check file too long for the usefulness judge (4 cases).
-
Test run started, each case with and without the skill.
-
-
8 Quality test, all skills together Running
Needs skill 3 (blog-draft).
-
New rules from the owner: one check plus at most one re-check; sources confirmed in the first pass by small parallel agents; lines with no first-hand proof are a warning the owner can ok; every reply short and in plain words.
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
Checklist due on the morning of 3 Nov.
No updates yet.
-
10 Board final yes, then main Skipped or out of order, not completed
-
On main since 01:00 on Alex's call. The Board's final yes is still pending.
-
blog-draft
In progress, 6 of 10 steps Takes an approved brief and writes a draft in the brand's voice, leaving marked gaps wherever first-hand material is missing. It checks its own draft with blog-check's tools.
Totals appear when the skill ships
6 of 10 steps done · 60%
-
1 Research (Scout) Done
-
Scout's research for blog-draft saved.
-
-
2 Plan (Rook) Done
-
Rook finished the rework spec for blog-draft. It gives a verdict of ship, fix, block or undecided instead of a score, and it never hands back a partial draft.
-
Rook's build plan for blog-draft is ready.
-
-
3 Merge and check (Alex) Done
-
Alex accepted Rook's choices: no partial drafts and no 0-100 total score.
-
Alex checked the plan and merged Scout's research into it.
-
-
4 Board yes on plan Done
-
Board approved the plan.
-
-
5 Build (Kit), Windows and Linux tests Done
-
Kit built the blog-draft rework. 841 tests pass across the build. One hygiene check still fails.
-
Test clean-up: tests cut from 895 to 573. Each test kept now catches a real planted bug. This work found one real gap in the brief guard; fixed and re-checked (17:40).
-
Build complete: four sessions, 714 automatic tests passing, Windows-style files covered.
-
Build session 3 done: 12 test cases written, 704 automatic tests passing. Final build session started.
-
Build sessions 1 and 2 done: the guard script, notes set-up and merge are in, with 672 tests passing. Session 3 (test cases) is running.
-
Kit started build session 1 of about 4: the draft guard script, its tests, and the skill skeleton.
-
-
6 Review (Sable) Done
-
Review complete after four rounds; every attack found is blocked or reported.
-
Review found 12 issues: ways a report could look checked when it was not, hidden or altered experience gaps, and Windows file-encoding breaks. Fixing before any paid test.
-
Independent review by Sable started.
-
-
7 Quality test, skill alone Running
-
First trial draft on a real topic was too thin: 1,376 words against 2,370 in the brief, with placeholders for visuals and calls to action. The quality judge did not finish its verdict, so that run is incomplete. A judge-only re-run and a full-length redraft with a review step come next.
-
Fix rounds (20:50 to 21:55): 16 of 17 pass. One case stays borderline: the judge finds the short draft thin. Paid re-runs stopped; the Board's own read of a sample draft is the bar.
-
Round 2 (04:22): 12 of 17. One real fault: the skill could edit an approved brief. Fixed with a guard that blocks those writes. Round 3 (08:15): 13 of 17. Round 4 (09:50): the full chain, brief to draft to check, passes. Four cases still fail on the honesty and ready-to-send judges. Rubric wording fixed (17:30). Re-run on 4 Nov.
-
First quality test: 12 of 17 cases passed, clearly better than without the skill. Found one real fault (the approved brief could be edited before the gate) and three honesty slips. All fixed; re-test next.
-
-
8 Quality test, all skills together Running
The chain: brief, then draft, then check.
-
Before writing, the skill now asks once: which voice, humaniser on or off, and any extra context. The answers are saved word for word.
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
No updates yet.
-
10 Board final yes, then main Not started
No updates yet.
blog-refresh
In progress, 6 of 10 steps An old post becomes a fact-by-fact update.
Totals appear when the skill ships
6 of 10 steps done · 60%
-
1 Research (Scout) Done
-
Scout's research saved, 29 sources (14:20).
-
-
2 Plan (Rook) Done
-
Rook finished the rework spec for blog-refresh.
-
Rook's build plan ready: 13 quality test cases, about 3 build sessions (14:30).
-
-
3 Merge and check (Alex) Done
-
Alex checked the plan and approved two small changes to blog-check (14:30).
-
-
4 Board yes on plan Done
-
Board approved the plan (20:10).
-
-
5 Build (Kit), Windows and Linux tests Done
-
All three build sessions done (20:40 to 20:50), plus a fix for unusual HTML files. 617 tests pass.
-
Kit started build session 1 (20:10).
-
-
6 Review (Sable) Not started
No updates yet.
-
7 Quality test, skill alone Done
-
Round 1 (21:10): 11 of 13. One fault in the skill (it did not read saved pages) and one in a test.
-
Kit fixed both. Re-run (21:35): 13 of 13 pass.
-
-
8 Quality test, all skills together Running
-
A light refresh can now be marked Ready to publish. Before, it always needed a brief and new first-hand proof, so it never could.
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
No updates yet.
-
10 Board final yes, then main Not started
No updates yet.
blog-plan
In progress, 5 of 10 steps Next month's topics, each tied to a goal.
Totals appear when the skill ships
5 of 10 steps done · 50%
-
1 Research (Scout) Done
-
Scout's research saved (20:12 to 20:25).
-
-
2 Plan (Rook) Done
-
Rook finished the rework spec for blog-plan. There is no single weighted priority number.
-
Rook's build plan ready (20:25). Board later asked for keyword data from a search data service; added to the plan (21:15).
-
-
3 Merge and check (Alex) Done
-
Alex accepted Rook's choice of no weighted priority number.
-
Alex checked the plan and approved small changes to two other skills (20:25).
-
-
4 Board yes on plan Done
-
Board approved the plan (20:50).
-
-
5 Build (Kit), Windows and Linux tests Done
-
Build complete. Fixes from the first test round added. 678 tests pass.
-
All three build sessions done (22:00 to 22:40). 672 tests pass on Linux.
-
-
6 Review (Sable) Not started
No updates yet.
-
7 Quality test, skill alone Running
-
Two fix rounds: 14 of 15 pass. One case stays borderline: the judge wants a stronger reason for one topic. Alex read the plan and it is grounded. Paid re-runs stopped.
-
Round 1 (00:20): 6 of 15 pass, 0.41 better than no skill. Two cases did not run because of a test set-up error. Kit is fixing.
-
-
8 Quality test, all skills together Running
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
No updates yet.
-
10 Board final yes, then main Not started
No updates yet.
blog-repurpose
In progress, 6 of 10 steps A post becomes LinkedIn and newsletter drafts, with no new claims.
Totals appear when the skill ships
6 of 10 steps done · 60%
-
1 Research (Scout) Done
-
Scout's research saved, 34 sources (00:30).
-
-
2 Plan (Rook) Done
-
Rook finished the rework spec for blog-repurpose. A post with a fix verdict can still be repurposed, minus any claims about its open faults.
-
Rook's build plan ready.
-
Rook started the build plan (00:30).
-
-
3 Merge and check (Alex) Done
-
Alex accepted Rook's choice that a fix verdict still allows repurposing, minus claims about open faults.
-
Alex checked the plan.
-
-
4 Board yes on plan Done
-
Board approved the plan with two changes: links as input, and where the link goes in a post.
-
-
5 Build (Kit), Windows and Linux tests Done
-
Both build sessions done. 713 tests pass on Linux.
-
-
6 Review (Sable) Not started
No updates yet.
-
7 Quality test, skill alone Done
-
Round 1: 10 of 14. Vera, our new tester, ran it.
-
Two fix rounds: 14 of 14 pass.
-
-
8 Quality test, all skills together Running
-
First two runs of the all-skills test: 1 of 4 cases passed, then 3 of 4. One case failed because its test brief promised material it did not supply, which is a fault in the test. Text planted in a source page did not leak into the output; the run without the skill did leak it.
-
The faulty all-skills test case is fixed. Its brief no longer promises material the case does not give.
-
Vera wrote 4 test cases that pass work from one skill to the next. A rules audit of all six skills was done first, and Kit fixed every must-fix item. The paid test run has started.
-
-
9 Human Windows check Not started
No updates yet.
-
10 Board final yes, then main Not started
No updates yet.